Send
Send
Send
123
Communications
in Computer and Information Science 601
Commenced Publication in 2007
Founding and Former Series Editors:
Alfredo Cuzzocrea, Dominik Ślęzak, and Xiaokang Yang
Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Phoebe Chen
La Trobe University, Melbourne, Australia
Xiaoyong Du
Renmin University of China, Beijing, China
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Orhun Kara
TÜBİTAK BİLGEM and Middle East Technical University, Ankara, Turkey
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Ting Liu
Harbin Institute of Technology (HIT), Harbin, China
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
More information about this series at http://www.springer.com/series/7899
Vladimir Vishnevsky Dmitry Kozyrev (Eds.)
•
123
Editors
Vladimir Vishnevsky Dmitry Kozyrev
V.A. Trapeznikov Institute of Control V.A. Trapeznikov Institute of Control
Sciences Sciences
Russian Academy of Sciences Russian Academy of Sciences
Moscow Moscow
Russia Russia
This volume contains a collection of revised selected full-text papers presented at the
18th International Conference on Distributed Computer and Communication Networks
(DCCN-2015), held in Moscow, Russia, October 19–22, 2015.
The conference is a continuation of traditional international conferences of the
DCCN series, which took place in Bulgaria (Sofia, 1995, 2005, 2006, 2008, 2009,
2014), Israel (Tel Aviv, 1996, 1997, 1999, 2001), and Russia (Moscow, 1998, 2000,
2003, 2007, 2010, 2011, 2013) in the last 18 years. The main idea of the conference is
to provide a platform and forum for researchers and developers from academia and
industry from various countries working in the area of theory and applications of
distributed computer and communication networks, to exchange their expertise, and to
discuss the perspectives of development and collaboration in this area. The content of
this volume is related to the following subjects:
1. Computer and communication networks architecture optimization
2. Control in computer and communication networks
3. Performance and QoS evaluation in wireless networks
4. Modeling and simulation of network protocols
5. Queueing and reliability theory
6. Wireless IEEE 802.11, IEEE 802.15, IEEE 802.16, and UMTS (LTE) networks
7. FRID technology and its application in intellectual transportation networks
8. Protocols design (MAC, Routing) for centimeter and millimeter wave mesh
networks
9. Internet and Web applications and services
10. Application integration in distributed information systems
11. Big data in communication networks
The DCCN 2015 conference received 126 submissions from 164 authors in 16
different countries. From these, 94 submissions were accepted and presented at the
conference, 38 of which were recommended by session chairs and selected by the
Program Committee for the proceedings, yielding an overall acceptance rate of 40 %.
All the papers selected for the proceedings are given in the form presented by the
authors. These papers are of interest to everyone working in the field of computer and
communication networks.
We thank all the authors for their interest in DCCN, the members of the Program
Committee for their contributions, and the reviewers for their peer-reviewing efforts.
DCCN 2015 was organized by the Russian Academy of Sciences (RAS), V.A.
Trapeznikov Institute of Control Sciences of RAS, the Research and Development
Company Information and Networking Technologies, and the Institute of Information
and Communication Technologies (Bulgarian Academy of Sciences).
Executive Committee
General Co-chairs
S.N. Vasilyev ICS RAS, Russia
V.M. Vishnevsky ICS RAS, Russia
Steering Committee
V.V. Rykov Gubkin Russian State University of Oil and Gas, Russia
D.A. Aminev ICS RAS, Russia
T. Atanasova IICT-BAS, Bulgaria
S.N. Kupriyahina ICS RAS, Russia
Publicity Chair
D.V. Kozyrev ICS RAS, Russia
Highway System (AHS) research of the 1990s. The mandate was issue by the
Intermodal Surface Transportation Efficiency Act. The goal was to have a fully
automated roadway or test track in operation by 1997. In 1997 twenty AHS-
equipped vehicles demonstrating hands and feet off driving on I-15 in San Diego,
California. U.S. Department of Transportation introduced the Intelligent Vehicle
Initiative in 1997. It was authorized in the 1998 as Transportation Equity Act
for 21st Century. In November 2003 it was announced allocating 75 MHz of spec-
trum at 5.9 GHz for research purposes to improve transportation mobility. The
basic concept of operations was that V2V and vehicle-to- infrastructure (V2I)
communication could support safety and mobility applications.
• Vehicle speed
• Vehicle position and heading (direction of travel)
• On or off the throttle (accelerating, driving, slowing)
• Brakes on, anti-lock braking
• Lane changes
• Stability control, traction control engaged
• Windshield wipers on, defroster on, headlamps on in daytime (raining, snowing)
• Brakes on, anti-lock braking
• Gear position (a car in reverse might be backing out of a parking stall)
In V2V network, every car, smart traffic signal could send, capture and
retransmit signals. Five to ten hops on the network would gather traffic condi-
tions a mile ahead. V2V warnings might come as different signals or cars brake
and steer around hazards. Eight automakers: GM, Ford, Toyota, Hyundai/Kia,
Honda, Volkswagen/Audi, Mercedes-Benz and Nissan/Infiniti, and 2500 vehicles
been taking part in a University of Michigan V2V project along 73 lane-miles
of roadway in Ann Arbor. GM suggests V2V will be effective when a quarter of
cars on the road are equipped.
Some projects and research in the area are listed below.
AutoNet2030 (co-operative Systems in Support of Networked Automated
Driving by 2030) is a European project connecting two domains of intensive
research: cooperative systems for Intelligent Transportation Systems and Auto-
mated Driving. The research issues re as follows: how can all these vehicles with
different capabilities most efficiently cooperate to increase safety and fluidity
of the traffic system? What kind of information should be exchanged? Which
organization (e.g. centralized or distributed) is the best? [2,4]
[7] proposed model which can describe the network-wide spatiotemporal
propagation of information while factoring the constraints arising from traf-
fic flow dynamics and V2V communications and with the integrated multi-layer
6 P. Stanchev and J. Geske
framework. The proposed model can describe the interdependencies among infor-
mation flow, traffic flow and V2V communication events by simultaneously track-
ing the dynamics of information flow and traffic flow. The captures information
flow propagation wave that can characterize how the density, speed, and loca-
tions of the vehicles lead to the dynamics of information flow. [3] addresses the
design, sensing, decision making, and acting infrastructure and several experi-
mental tests that have been carried out to evaluate both platforms and proposed
algorithms. The communication and security aspects are also investigated.
In [1] multimedia-based ad-hoc networking (VANET) is tested to fulfill
demands in a vehicular environment, and the need to evaluate the current stan-
dards. In [6] set of algorithms that determine the crossing order are fed with
information about surrounding vehicles: actual and further GPS position, speed
and an identification number. In [8] a novel distributed intrusion detection sys-
tem designed for a vehicular ad hoc network by combining static and dynamic
detection agents that can be mounted on central vehicles, and a control cen-
ter where the alarms about possible attacks on the system are communicated.
The proposed DIDS can be used in both urban and highway environments for
real time anomaly detection with good accuracy and response time. In [9] a
vision-based multi-object tracking system for checking the plausibility of V2V
communication is presented. The system is addressing the challenge of fusing
relative sensor observations as provided by a Mobil Eye vision-system with time-
delayed absolute GNSS-based measurements from Cooperative Awareness Mes-
sages (CAMs) as provided by V2V.
[11] evaluate the effects of buildings on the vehicle-to-vehicle performance at
urban intersections based on a profound simulation campaign. Due to the two
dimensional nature of intersection topologies, we investigate the performance
of V2V communication by analyzing packet delivery ratios and packet drop
rates with respect to sender and receiver’s position under varying node density
and intersection layout. [12] provide an overview of ITS activities and status
worldwide. It is divided into 4 main parts: standards/specifications applicable
for ITS, spectrum allocation and channel plans, 802.11p PHY details, and test
and measurement solutions to aid in design and verification of ITS devices and
systems.
The purpose of [5] is to assess the readiness for application of V2V communi-
cations, a system designed to transmit basic safety information between vehicles
to facilitate warnings to drivers concerning impending crashes. The United States
Department of Transportation and NHTSA have been conducting research on
this technology for more than a decade. A suggested V2V Security System Design
for Deployment and Operations diagram in the report is giving in Fig. 1. The
schema presents a full deployment model. This diagram shows also in dotted
lines the initial deployment model where there is no Intermediate CA (certifi-
cate authority) and the Root CA talks to the MA (misbehavior authority), PCA
(pseudonym certificate authority), and ECA (enrollment certificate authority).
SCMS stands for Security Credentials Management System and CRL stands for
certificate revocation list.
Autonomous Cars. History. State of Art. Research Problems 7
Further research to move toward deployment has been identified and will be
conducted to address the following:
4 Conclusions
Further tests and developments have to be conducted for the V2V applications
and combined with the sensor-based systems. The current test procedures should
be modified to reflect a greater range of speeds and a greater variety of road
geometry configurations. Research in V2V safety has to identify the performance
of V2V safety applications. Many questions have to be answered.
References
1. Akbar, M., Khan, M., Khaliq, K., Qayyum, A., Yousaf, M.: Evaluation of IEEE
802.11n for multimedia application in VANET, science direct. Procedia Comput.
Sci. 32, 953–958 (2014)
2. EU 7th Framework Programme, Project number 610542, Specifications for the
enhancement to existing LDM and cooperative communication protocol standards.
http://www.autonet2030.eu/wp-content/uploads/2015/02/D3.2-Specifications-
cooperative-communication-protocol-standards-draft-for-approval.pdf
3. Fernandes, L., Souza, J., Pessin, G., Shinzato, P., Sales, D., Mendes, C., Prado, M.,
Klaser, R., Magalhaes, A., Hata, A., Pigatto, D., Branco, K., Grassi Jr., V., Osorio,
F., Wolf, D.: CaRINA intelligent robotic car: architectural design and applications.
J. Syst. Archit. 60(4), 372–392 (2013)
4. Fortelle, A., Qian, X., Diemer, S., Grégoire, J., Moutarde, F., Bonnabel, S.,
Marjovi, A., Martinoli, A., Llatser, I., Festag, A., Katrin, S.: Network of Auto-
mated Vehicles: The Autonet2030 Vision, ITS World Congress, Detroit (2014)
5. Harding, J., Powell, G., R., Yoon, R., Fikentscher, J., Doyle, C., Sade, D., Lukuc,
M., Simons, J., Wang, J.: Vehicle-to-vehicle communications: Readiness of V2V
technology for application. (Report No. DOT HS 812 014). National Highway Traf-
fic Safety Administration, Washington, DC, August 2014
6. Hernandez-Jayo, U., De-la-Iglesia, I., Lacoume, I., Mammu, A.: I-CROSS: intersec-
tion crossing warning application based on V2V communications. In: IEEE Vehic-
ular Networking Conference, pp. 206–209 (2013)
7. Kim, Y., Peeta, S., He, X.: An analytical model to characterize the spatiotemporal
propagation of information under vehicle-to-vehicle communications,. In: IEEE
17th International Conference on Intelligent Transportation Systems (ITSC), 8–11
October 2014, Qingdao, pp. 1142–1147 (2014)
8. Maglaras, A.: A novel distributed intrusion detection system for vehicular Ad Hoc
networks. (IJACSA) Int. J. Adv. Comput. Sci. Appl. 6(4), 101–106 (2015)
9. Obst, M., Hobert, L., Reisdorf, P.: Multi-sensor data fusion for checking plausi-
bility of V2V communications by vision-based multiple-object tracking. In: IEEE
Vehicular Networking Conference, pp. 143–150 (2014)
10 P. Stanchev and J. Geske
1 Introduction
Our paper deals with mobile web presentations of location-based services. How
can we present some local (attached to a certain geographical location) informa-
tion to mobile users? We are talking about programming (creating) mobile web
sites, which content pages correspond to the current location of the mobile user.
The traditional scheme is very straightforward. We have to determine the user’s
location and then create a dynamic web page, the issuance of which is clearly
defined by specific geographical coordinates. For example, geo-location is a part
of HTML5 standard [1].
As soon as web application obtained (as per user permission, of course) geo-
coordinates, it can build a dynamic web page, which content depends on the
current location (content is associated with obtained location). Technically, we
can render our dynamic page on the client side (right in the browser), when
application requests data from server via some asynchronous calls (AJAX) [2],
or right on our server (in some CGI-script). In both cases obtained location
info is used as a parameter either to AJAX script or to CGI script. For some
of the applications (classes of applications), we may use several location-related
datasets (e.g., so-called geo-fence [3]), but the common principles are similar. It
is so called Location Based Services (LBS) [4].
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 11–18, 2016.
DOI: 10.1007/978-3-319-30843-2 2
12 D. Namiot and M. Sneps-Sneppe
There are different methods for obtaining location information for mobile
users [5]. Not all of them use GPS (GLONAS) positioning actually. Alterna-
tive approaches use Wi-Fi, Cell ID, collaborative location, etc. [6]. The above-
mentioned geo-location in HTML5 has been a wrapper (interface) for location
service. For the most of LBS, their top-level architecture is standard. LBS use
obtained location info as a key for any database (data store) with location-
dependent data. Location info is actually no more than a key for linking physical
space (location) and virtual (e.g., coupon for the store). Only a small number
of services actually use the coordinates. The typical example is indoor location
based services. The paradigm Location first requires a digital map for an indoor
space. This map should be created prior to the deployment, and it should be sup-
ported in an actual state during service’s life time. On the other hand, there is
a direction, called context-aware computing. In context-aware computing (ubiq-
uitous computing) services can use other information (not related to geographic
coordinates) as the “characteristics” of a user’s location. Simplistically, the con-
text is any additional information on the geographical location [7,8]. In this case,
additional information (context), with the presence of certain metrics can serve
as a unique (up to a certain approximation, of course) feature of a user’s location.
Or, in other words, we can substitute geo-location with context identification.
Why might it be necessary? The typical example is indoor LBS [9]. Traditional
geo-positioning can be difficult and positioning accuracy may be insufficient to
distinguish the position of the mobile subscriber within the same premises. And
yet, it is the distinction between positions within the same space (buildings) may
be important for all kinds of services (for example, the buyer is located on the
first or second floor of the hall).
Actually, it is a starting point for new approaches in LBS architecture, when
the stage with obtaining (detecting) location info could be completely eliminated.
Indeed, if location info is no more than a key for some database, then why do
not replace geo-keys (e.g., latitude and longitude) with context-related IDs? It
is sufficient to identify context and use this identification to search data.
The rest of the paper is organized as follows. In Sect. 2, we describe context
identification. In Sect. 3, we describe how this identification could be used in web
programming. In Sect. 4, we discuss the generic approaches for incorporating
sensing information into web pages.
2 Network Proximity
One of the widely used methods for the identification of context is the use of wire-
less network interfaces of mobile devices (Wi-Fi, Bluetooth). The reasons for this
are straightforward. On the first hand, these interfaces are supported in all mod-
ern smart-phones. Secondly, for obvious reasons, monitoring of network interfaces
is directly supported and executed by the mobile operating systems. Therefore,
a survey of network interfaces on the application level can be simplified and not
cause additional power consumption, as compared with, for example, a specially
organized monitoring for the accelerometer. Information received through the
On Hyper-local Web Pages 13
network interface is used to estimate the proximity of the mobile user to the ele-
ments of the network infrastructure (network proximity [10]). Note, that other
mobile devices can act as these elements too (e.g., Wi-Fi access point, opened
right onto mobile phone [11]). The classical form for collecting data about Wi-Fi
devices are so-called Wi-Fi fingerprints sets [12]. Wi-Fi fingerprints are digital
objects that describe availability (visibility) for network nodes. Their primary
usage is navigation related tasks. The alternative approach lets users directly
associate some data chunks with existing (or artificially created) network nodes.
In other words, it is a set of user generated links between network nodes and
some content that could be used by those in proximity to networks nodes. This
approach is presented in SpotEx project and associated tools [13,14]. SpotEx lets
users create a set of rules (logical productions) for linking network elements and
available content. A special mobile application (context-aware browser) is based
on the external set of rules (productions, if-then operators). The conditional part
of the each rule includes predicates with the following objects:
– identity for Wi-Fi network (name, MAC-address)
– RSSI (signal strength),
– time of the day (optionally),
In other words, it is a set of operators like this:
IF AccessPointIsVisible (‘Cafe’) THEN { show content for Cafe }
Block { show content for Cafe } is some data (information) snippet presented
in the rule. Each snippet has got a title (text) and some HTML content (it
could be simply a link to any external site for example). Snippets could present
coupons/discount info for malls, news data for campuses, etc. The context-aware
browser (mobile application) maps current network environment against existing
database, detects relevant rules (fires them) and builds a dynamic web page. This
web page is presented to a mobile user in proximity. In fact, even the name of the
application (context-aware browser) suggests the movement of this functionality
in a mobile browser. This would eliminate the separate rule base as well as the
special (separate) application. In fact, the standard mobile browser should play
a role of this application. Rules for the content (data snippets) must be specified
directly on the mobile web pages. And data snippets itself are HTML code chunks
anyway. As applied implementations, we can mention, for example, Internet of
Things applications [15,16]. The usage is very transparent. Data snippets (data,
presented to mobile users) depends on visibility for some Wi-Fi access points.
It lets us specify the positions for mobile users inside of some building (campus,
etc.) Mobile users will see different information for different positions. And this
approach does not use geo-coordinates at all. The next interesting direction is
EU project FI-WARE [17]. Integration with the FI-CONTENT platform is one
of the nearest goals.
3 Information Services
Technically, for the reuse of information about network proximity, we can talk
about the two approaches. At the first hand, the implementation of a mobile
14 D. Namiot and M. Sneps-Sneppe
browser can follow the same ideology that supports geo-coding in HTML5 [18].
A function from browser’s interface
navigator.geolocation.getCurrentPosition()
accepts as a parameter some user-defined callback (another function). The call-
back should be called as soon as geo-location is completed. Obtained data should
be passed as parameters. Note, that the whole process is asynchronous. By the
analogue with the above-mentioned model, a mobile browser can add a new
interface function. E.g., getNetworks() this function will accept a user-defined
callback for accumulating network information (current fingerprint). A good can-
didate for data model is JSON. The browser will pass fingerprint as a JSON array
to a user-defined callback. Each element from this array describes one network
and contains the following information:
This approach lets us proceed network proximity right in JavaScript (in other
words, right on the web page). Actually, by the similar manner we can work with
other sensors too. It is so-called Data Program Interface [26]. We would like to
see something similar as a standard feature in the upcoming versions of Android.
5 Conclusion
The paper discusses the use of information about the network environment to
create dynamic web pages. We propose several approaches to the implementation
of a mobile browser that can handle data on a network (network proximity) to
provide users with information tied to the current context. Also, we considered
possible implementation details. The basic idea is to separate the functional for
scanning network information and real data exchange.
On Hyper-local Web Pages 17
References
1. Holdener, A.T.: HTML5 Geolocation. O’Reilly Media Inc., Sebastopol (2011)
2. Namiot, D., Sneps-Sneppe, M.: Where are they now – safe location sharing. In:
Andreev, S., Balandin, S., Koucheryavy, Y. (eds.) NEW2AN/ruSMART 2012.
LNCS, vol. 7469, pp. 63–74. Springer, Heidelberg (2012)
3. Namiot, D., Sneps-Sneppe, M.: Geofence and network proximity. In: Balandin, S.,
Andreev, S., Koucheryavy, Y. (eds.) NEW2AN/ruSMART 2013. LNCS, vol. 8121,
pp. 117–127. Springer, Heidelberg (2013)
4. Prasad, M.: Location Based Services, GIS Development, pp. 3–35 (2002)
5. Tabbane, S.: Location management methods for third generation mobile systems.
IEEE Commun. Mag. 35(8), 72–78 (1997). 10.1109/35.606034
6. Namiot, D.: Context-aware browsing - a practical approach. In: 6th International
Conference on Next Generation Mobile Applications, Services and Technologies
(NGMAST), pp. 18–23 (2012). http://dx.doi.org/10.1109/NGMAST.2012.13
7. Schilit, G., Theimer, B.: Disseminating active map information to mobile hosts.
IEEE Netw. 8(5), 22–32 (1994). 10.1109/65.313011
8. Namiot, D., Sneps-Sneppe, M.: Context-aware data discovery. In: 2012 16th Inter-
national Conference on Intelligence in Next Generation Networks (ICIN), pp. 134–
141 (2012). http://dx.doi.org/10.1109/ICIN.2012.6376016
9. Kolodziej, K., Danado, J.: In-building positioning: modeling location for indoor
world. In : Proceedings of the 15th International Workshop on Database and
Expert Systems Applications, pp. 830–834. IEEE (2004). http://dx.doi.org/10.
1109/DEXA.2004.1333579
10. Sharma, P., Xu, Z., Banerjee, S., Lee, S.: Estimating network proximity
and latency. ACM SIGCOMM Comput. Commun. Rev. 36(3), 39–50 (2006).
10.1145/1140086.1140092
11. Namiot, D.: Network proximity on practice: context-aware applications and Wi-Fi
proximity. Int. J. Open Inf. Technol. 1(3), 1–4 (2013)
12. Cheng, Y.C., Chawathe, Y., LaMarca, A., Krumm, J.: Accuracy characterization
for metropolitan-scale Wi-Fi localization. In: Proceedings of the 3rd International
Conference on Mobile Systems, Applications, and Services, pp. 233–245. ACM
(2005). http://dx.doi.org/10.1145/1067170.1067195
13. Namiot, D., Schneps-Schneppe, M.: About Location-aware Mobile Messages:
Expert System Based on WiFi Spots. In: 5th International Conference on Next
Generation Mobile Applications, Services and Technologies (NGMAST), pp. 48–
53. IEEE (2011). http://dx.doi.org/10.1109/NGMAST.2011.19
14. Namiot, D., Sneps-Sneppe, M.: Wi-Fi proximity as a service. In: 1st Interna-
tional Conference on Smart Systems, Devices and Technologies, SMART, pp. 62–68
(2012)
15. Schneps-Schneppe, M., Namiot, D.: Open API for M2M applications: what is next?.
In: 8th Advanced International Conference on Telecommunications, AICT, pp. 18–
23 (2012)
16. Namiot, D., Schneps-Schneppe, M.: Smart cities software from the developer’s
point of view. In: 6th International Conference on Applied Information and Com-
munication Technologies (AICT), LUA Jelgava, Latvia, pp. 230–237 (2013)
17. Castrucci, M., Cecchi, M., Priscoli, F.D., Fogliati, L., Garino, P., Suraci, V.: Key
concepts for the future internet architecture. In: Future Network & Mobile Summit
(FutureNetw), pp. 1–10. IEEE (2011). http://dx.doi.org/10.1145/1067170.1067195
18 D. Namiot and M. Sneps-Sneppe
18. Aghaee, S., Cesare, P.: Mashup development with HTML5. In: Proceedings of the
3rd and 4th International Workshop on Web APIs and Services Mashups, pp. 10.
ACM (2010). http://dx.doi.org/10.1145/1944999.1945009
19. Amatya, S., Kurti, A.: Cross-platform mobile development: challenges and oppor-
tunities. In: Trajkovik, V., Anastas, M. (eds.) ICT Innovations 2013. Advances in
Intelligent Systems and Computing, vol. 231, pp. 219–229. Springer, Heidelberg
(2014)
20. Paul, A., Steglich, S.: Virtualizing devices. In: Crespi, N., Magedanz, T., Bertin,
Emmanuel (eds.) Evolution of Telecommunication Services. LNCS, vol. 7768, pp.
182–202. Springer, Heidelberg (2013)
21. Cazenave, F., Quint, V., Roisin, C.: Timesheets. js: when SMIL meets HTML5 and
CSS3. In: Proceedings of the 11th ACM Symposium on Document Engineering,
pp. 43–52. ACM (2011). http://dx.doi.org/10.1145/2034691.2034700
22. Zheng, C., Shen, W., Ghenniwa, H.H.: An intents-based approach for service dis-
covery and integration. In: IEEE 17th International Conference on Computer Sup-
ported Cooperative Work in Design (CSCWD), pp. 207–212. IEEE (2013). http://
dx.doi.org/10.1109/CSCWD.2013.6580964
23. Chin, E., Felt, A.P., Greenwood, K., Wagner, D.: Analyzing inter-application com-
munication in android. In: Proceedings of the 9th International Conference on
Mobile Systems, Applications, and Services, pp. 239–252. ACM (2011). http://dx.
doi.org/10.1145/1999995.2000018
24. Firtman, M.: Programming the Mobile Web, pp. 619–624. O’Reilly Media Inc,
Sebastopol (2013)
25. Oliveira, L., Ribeiro, A.N., Campos, J.C.: The mobile context framework: providing
context to mobile applications. In: Streitz, N., Stephanidis, C. (eds.) DAPI 2013.
LNCS, vol. 8028, pp. 144–153. Springer, Heidelberg (2013)
26. Namiot, D., Sneps-Sneppe, M.: On software standards for smart cities: API or DPI.
In: Proceedings of the 2014 ITU Kaleidoscope Academic Conference: Living in a
Converged World-Impossible Without Standards?, pp. 169–174. IEEE (2014)
Analysis of Two-Server Queueing Model
with Phase-Type Service Time Distribution
and Common Phases of Service
1 Introduction
The simplest queueing models suggest that customer’s service time is exponen-
tially distributed. This allows to avoid the necessity of taking into account the
elapsed service time when the Markovian process describing behavior of the
queue should be constructed. It is quite obvious that in many real world appli-
cations of queueing theory assumption about the exponential distribution does
not hold true and more general distributions of service time have to be consid-
ered. If service time has an arbitrary distribution, it is mandatory to take into
account the elapsed or residual service times, e.g., by introducing supplementary
variables or considering the embedded Markov chains. This may lead to huge
analytical difficulties in analysis of the Markovian process describing dynamics
of the queue, especially when multi-server queue is under study. To avoid such
difficulties, so called phase type (P H) distribution, as natural extension of previ-
ously well-known Erlangian and hyper-exponential distribution, was offered, see,
e.g., [1]. Good property of this distribution is its generality. Generally speaking,
any distribution can be approximated, in sense of a weak convergence, by the
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 19–29, 2016.
DOI: 10.1007/978-3-319-30843-2 3
20 C. Kim et al.
P H type distribution, see, e.g., [3]. Random time having the P H type distribu-
tion can be interpreted as the sequence of phases duration of each of which has
exponential distribution. The total number of existing phases of service is finite.
However, implementation of the phases during the service of a customer may
be repeated random number of times. So, in concrete realization of the random
time having the P H type distribution the number of phases in a sequence is
random.
Formal definition of P H type distribution is given in the next section. For
purposes of this paper, we slightly rephrase this definition as follows. There is
a virtual network with nodes (phases), say, {1, . . . , M }. Random time having
the P H type distribution with an irreducible representation (β, S) is the time
during which some virtual customer stays in this network conditional on the
fact that this virtual customer starts its staying in the network from the visit
to the state m of the network with probability β m , m = 1, M , then it makes
transitions inside of the set {1, . . . , M } with intensities given by the entries of
the matrix S or leaves the network from any state m with intensity, which is
the mth component of column vector S0 = −Se, where e denotes unit column
vector. In brief, as it was already noted above, random time having the P H
type distribution consists of the random number of virtual phases, duration of
which is exponentially distributed. This allows to replace keeping track of the
continuous elapsed or residual service time by the keeping track of the discrete
current phase of the service what greatly simplifies the analysis.
So, phases of service in definition of P H type distribution may be just the
virtual entities. However, in many real world situations, random time having
P H type distribution may represent the sequence of real phases. E.g., process-
ing time of the query in data base consists of implementation of a sequence of
input/output operations alternating with the use of CPU. Processing of a car
in service station consists of a sequence of technological operations. Security
control in airport includes screening of a luggage and passengers with possible
additional personal individual passenger inspection. If the service is provided by
the single server, no collisions occur. But, if there are several servers operating
in parallel, collisions may occur because the servers use some common resources,
e.g. memory and CPU, tables and indices of data base, different equipment of
a car service station, screening devices and cabins for the personal individual
inspection, etc.
Traditionally in analysis of multi-server queues with P H type service time
distribution, it is assumed that service processes in the servers are independent.
To the best of our knowledge, the systems with interference of phases of service in
different servers are not considered in literature. In this paper, we start research
of such systems from a relatively simple model where there are only two servers,
state spaces of the virtual networks, in terms of which P H type service time
distribution is interpreted, coincide and if some phase of the service by a server
is required while this phase is busy by another server this phase of service is
postponed until it will be released by another server. In some sense, in this model
we somehow unite the problems considered in two different popular branches
Analysis of Two-Server Queueing Model with Phase-Type Service 21
2 Mathematical Model
Queueing system with two servers and a buffer of infinite capacity is considered.
The structure of the system under study is presented in Fig. 1.
buffer
to the server that currently occupies the phase under the conflict. Number 2 is
appointed to the server that currently waits for releasing this phase.
The customers from the buffer are impatient, i.e., the customer leaves the
buffer and the system after an exponentially distributed waiting time described
by the parameter α, 0 < α < ∞.
It is easy to see that the dynamics of the system under study is described by the
following regular irreducible multi-dimensional continuous-time Markov chain
ξt = {it , rt , νt , nt , mt }, t ≥ 0,
{i, 0, ν, n, m}, i ≥ 2, n = 1, M , m = 1, M , m = n
{i, 1, ν, n}, i ≥ 2, n = 1, M , ν = 0, W .
For further use throughout this paper, we introduce the following notation:
⎛ ⎞
0T S1,2 e1,2 . . . S1,M e1,M
⎜ S e 0T . . . S2,M e2,M ⎟
2,1 2,1
(0,1) ⎜ ⎟
Qi,i = IW̄ ⊗ ⎜ .. .. .. ⎟ + diag{a1 , . . . , aM } , i > 1,
⎝ . . ... . ⎠
SM,1 eM,1 SM,2 eM,2 ... 0T
⎛ ⎞
0 S1,2 c1,2 . . . S1,M c1,M
⎜ S2,1 c1,M 0 . . . S2,M c1,M ⎟
(1,0) ⎜ ⎟
Qi,i = IW̄ ⊗ ⎜ .. .. .. ⎟,
⎝ . . ... . ⎠
SM,1 c1,M SM,2 c1,M . . . 0
(1,1)
Qi,i = D0 ⊕ diag{S1,1 , . . . , SM,M } − (i − 2)αIW̄ M , i > 1,
Q0,1 = D1 ⊗ β,
Q1,2 = (0,0)
Q1,2 Q1,2
(0,1)
,
(0,0)
Q1,2 = D1 ⊗ diag{β 1 , . . . , β M },
(0,1)
Q1,2 = D1 ⊗ diag{β1 , . . . , βM },
D1 ⊗ IM (M −1) O
Q =
+
,
O D1 ⊗ IM
(0,0)
Q2,1
Q1,0 = IW̄ ⊗ S0 , Q2,1 = (1,0) , i > 1,
Q2,1
⎛ ⎞
S0 1 I1+
(0,0) ⎜ . ⎟
Q2,1 = IW̄ ⊗ ⎝ .. ⎠ + diag{(S0 ) , l = 1, M } ,
l
(S0 )M IM
+
⎛ ⎞
(S0 )1 ã1
(1,0) ⎜ .. ⎟
Q2,1 = IW̄ ⊗ ⎝ . ⎠,
(S0 )M ãM
(0,0) (0,1)
Qi,i−1 Qi,i−1
Qi,i−1 = (1,0) (1,1) , i > 2,
Qi,i−1 Qi,i−1
26 C. Kim et al.
⎛ ⎞
S0 1 B1
(0,0) ⎜ . ⎟
Qi,i−1 = IW̄ ⊗ ⎝ .. ⎠ +diag{(S0 ) β l , l = 1, M } +(i−2)αIW̄ M (M −1) ,
l
(S0 )M BM
⎛ ⎞
(S0 )1 C1
(0,1) ⎜ .. ⎟
Qi,i−1 = IW̄ ⊗ ⎝ . ⎠ + diag{S l
β
0 l , l = 1, M } ,
(S0 )M CM
(1,0)
Qi,i−1 = IW̄ ⊗ diag{(S0 )l β l , l = 1, M },
(1,1)
Qi,i−1 = IW̄ ⊗ diag{(S0 )l βl , l = 1, M } + (i − 2)αIW̄ M .
y(Y0 + Y1 + Y2 ) = y, ye = 1.
It is easy to verify that for the considered Markov chain the matrices Y0 , Y1
and Y2 have the following form:
Y0 = I, Y1 = O, Y2 = O.
Thus, the following limits (stationary probabilities) exist for any set of the
system parameters:
i ≥ 0, r = 0, 1, ν = 0, W , n = 1, M , m = 1, M .
Then let us form the row vectors π i of the stationary probabilities as follows:
(π 0 , π 1 , . . . )Q = 0, (π 0 , π 1 , . . . )e = 1. (2)
To solve system (2), we use the numerically stable algorithm for computation
of the probability vectors π i , i ≥ 0, developed in [10] which effectively uses
information about the asymptotic behavior of the Markov chain ξt , t ≥ 0, and
the sparse structure of the generator Q.
4 Performance Measures
The average number Nbuf f er of customers in the buffer is computed by
∞
Nbuf f er = (i − 2)π i e.
i=3
The probability Ploss that an arbitrary customer will be lost (due to impa-
tience) is computed by
αNbuf f er
Ploss = .
λ
The average intensity λout of flow of customers who receive service is com-
puted by
λout = λ(1 − Ploss ).
5 Conclusion
Two server queueing model with an infinite buffer and Markovian arrival process
is analysed. Service times by two servers have phase type distribution with coin-
ciding state spaces of underlying Markov chains. The phases of service times at
two servers are implemented independently if the underlying Markov chains of
services currently have different states (phases). If the required phases of ser-
vice coincide, service by one of the servers is postponed until the phase will be
released by the competitive server. Generator of multi-dimensional Markov chain
describing behavior of the system is written down. Formulas for computation of
the key performance measure of the system in terms of stationary probabilities
of the Markov chain are presented. It is planned to extend the results to the
cases when the number of servers is more than two, when the state spaces of
underlying processes of service coincide partially, when the vectors and subgen-
erators defining irreducible representations of service times may we selected in
such a way as to minimize possibility of conflicts, when the number of active
servers can be dynamically changed, etc.
References
1. Neuts, M.F.: Matrix-geometric Solutions in Stochastic Models. The Johns Hopkins
University Press, Baltimore (1981)
2. Bocharov, P.P., D’Apice, C., Pechinkin, A.V., Salerno, S.: Queueing Theory. VSP,
Utrecht-Boston (2004)
3. Asmussen, S.: Applied Probability and Queues. Springer, New York (2003)
4. Lucantoni, D.: New results on the single server queue with a batch Markovian
arrival process. Commun. Stat. Stoch. Models 7, 1–46 (1991)
Analysis of Two-Server Queueing Model with Phase-Type Service 29
5. Heyman, D.P., Lucantoni, D.: Modelling multiple IP traffic streams with rate lim-
its. IEEE/ACM Trans. Netw. 11, 948–958 (2003)
6. Klemm, A., Lindermann, C., Lohmann, M.: Modelling IP traffic using the batch
Markovian arrival process. Perform. Eval. 54, 149–173 (2003)
7. Chakravarthy, S.R.: The batch Markovian arrival process: a review and future
work. In: Krishnamoorthy, A., Raju, N., Ramaswami, V. (eds.) Advances in Prob-
ability Theory and Stochastic Processes, pp. 21–29. Notable Publications Inc.,
New Jersey (2001)
8. Graham, A.: Kronecker Products and Matrix Calculus with Applications. Ellis
Horwood, Cichester (1981)
9. Klimenok, V.I., Dudin, A.N.: Multi-dimensional asymptotically quasi-Toeplitz
Markov chains and their application in queueing theory. Queueing Syst. 54,
245–259 (2006)
10. Dudina, O., Kim, C., Dudin, S.: Retrial queueing system with Markovian arrival
flow and phase-type service time distribution. Comput. Ind. Eng. 66, 360–373
(2013)
Optimization of Topological Structure
of Broadband Wireless Networks Along
the Long Traffic Routes
1 Introduction
Broadband wireless networks and communication channels have become recently
one of the main directions of development of the telecommunications indus-
try [1,2]. The development of networks of this class for creation of a modern
infrastructure for multimedia data transmission along the long transport routes
is one of the major problems of implementation of new transport routes and
pipelines and exploiting the existing ones.
Creation of such communication infrastructure allows to: provide the operat-
ing control over the technical parameters of a route by the means of high-speed
data transfer from sensors and data units to the Control Center; provide the
security control over the route sections and strategically important objects using
data from the video surveillance systems; provide the voice communication (IP-
telephony) and transmission of multimedia information between the stationary
and mobile objects on long highways as well as communication with the Control
Center, etc.
Due to the high practical demands for high-performance communication net-
works along the long transport routes based on IEEE 802.11-2012 [3] hardware
and software, in recent years a considerable number of studies on this class of
This research was financially supported by the Ministry of Education and Sci-
ence of the Russian Federation in the framework of the applied research project
№14.613.21.0020 of 22.10.2014 (RFMEFI61314X0020).
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 30–39, 2016.
DOI: 10.1007/978-3-319-30843-2 4
Optimization of Topological Structure 31
networks have appeared [4–11]. Particularly, in [4] the problem of wireless net-
work base stations deployment, maximizing the network coverage is investigated
under constraints on the total network cost. The initial data for solution of the
problem are the potential locations for deployment of the base stations, and
the beforehand-collected statistics of traffic from fixed and mobile users. The
close problem of the deployment of base stations, maximizing the coverage, is
formulated in [5]. For the analytical description of the problem, it is modeled as
a Maximum Coverage with Time Threshold Problem (MCTTP), and a genetic
algorithm is used to solve it. Strategies for roadside units (RSUs) placement
based on the road traffic characteristics, aiming at improving connectivity in
vehicular ad hoc networks, are presented in [6]. To divide the coverage area of
each RSU, the authors propose an Expansion and Coloration Algorithm (ECA).
The average connectivity model for all vehicles in the network is established
based on the results obtained from ECA. The RSUs placement problem is for-
mulated as a combinatorial optimization problem, of which the objective is to
maximize the average connectivity probability by searching for an optimal posi-
tion combination of the given RSUs. Taking part of an actual urban road net-
work as an example, the RSUs placement problem is calculated and the optimal
placement scheme is evaluated. Simulation results show that the optimal place-
ment scheme obtained from the proposed strategy leads to the best connectivity
compared to uniform placement and hot-spot placement.
The problem of roadside units placement in IEEE 802.11p/WAVE (wire-
less access in vehicular environment) networks is studied in [7]. An analytical
model is presented that allows to analyze the delay of data transmission in com-
munication networks along highways. The cases of bound and unbound base
stations are investigated. It is shown that only those deployment strategies are
efficient in which roadside units are connected to each other within the line-of-
sight range. Paper [8] designs a connectivity analysis scheme for the roadside-
to- vehicle telematics network based on the real movement history of vehicle
objects collected from taxi telematics system currently in operation, aiming at
providing a useful guideline and information to build a telematics network. The
implemented analyzer can locate the current and previous positions of all vehi-
cles, decide whether it can be connected to an RSU, and calculate the duration
of disconnected state, taking into account the transmission range, the number
of RSUs, and RSU deployment. The RSU placement scheme can improve the
network coverage exploiting the analysis result.
The results of simulation and measurement of the performance of a roadside
unit placement scheme for the vehicular telematics network on the road network
of Jeju city (South Korea) are presented in [9]. The calculated optimal topol-
ogy of the backbone wireless network provides improvement of connectivity and
reduction of the disconnection interval for the given number of roadside units,
the transmission range, and the overlap ratio. A research problem of finding
the optimal locations to place dissemination points (i.e. roadside infrastructure
nodes for information dissemination) is considered in [10]. In this paper a novel
approach is proposed for dissemination points placement in grid road networks
32 V.M. Vishnevsky et al.
without knowing trajectories. Based on the analysis of path number between two
intersections, a probabilistic model is proposed to get the trajectories estimation
of vehicles. The problem of the roadside unit placement in vehicular networks is
studied in [11], where the authors focus on the highway-like scenario in which
there may be multiple lanes with exits or intersections along the road. In the
proposed model, each vehicle can access RSUs in two ways: (1) direct delivery,
which occurs when the vehicle is in the transmission range of the RSUs, and
(2) multi-hop relaying, which takes place when the vehicle is out of RSU trans-
mission range. Both access patterns in this placement strategy are worked out
and this placement problem is formulated via an integer linear programming
model such that the aggregate throughput in the network can be maximized.
The impact of wireless interference, vehicle population distribution, and vehi-
cle speeds are also taken into account in the formulation. The performance of
the proposed placement strategy is evaluated via NS-2 simulations to generate
vehicle mobility patterns.
In this article is considered RSU optimal placement problem (discrete Case)
with number of practical limitations. This limitations was founded while devel-
opment broadband wireless network for Ring Road of Kazan City (M7 “Volga”).
Let give some definitions. Let the section α has orientation and xA < xB .
Designate r(t), R(t), c(t) as coverage radius, transmission range and station cost
of station type t respectively.
M(Δ(P ))
δ(P ) = , (3)
M(α)
Considering all definitions above, the general formulation of the problem is:
with limitations
⎧
⎨P̂ − an extended correct arrangement
N
⎩ c(t(pi )) ≤ C, - total system cost is less than budget C
i=1
34 V.M. Vishnevsky et al.
v0,0 = 1,
(11)
vm+1,0 = 1,
Cost Limitations:
m+1
As it follows from 4 the sum θi is exactly total cost of set all installed
i=0
RSU.
Limitations of Coverage:
−
Condition 6. Coverage segments ρ+ i ρi must be less than coverage radius of
RSU in point i:
n
∀i = 0 . . . m + 1, r(t(j))vi,j ≥ ρ+
i ,
j=1
n
(15)
∀i = 0 . . . m + 1, r(t(j))vi,j ≥ ρ−
i ,
j=1
Ie real coverage segment is everywhere equal ar less than declared by RSU type.
−
Condition 7. Every segments ρ+
i ρj don’t have more than one mutual points:
−
∀i = 0 . . . m + 1, ∀j = 0 . . . m + 1, xi + ρ+
i ≤ xj − ρj , (16)
In terms of coverage area it means that the segments not intersect each other.
Limitations RSUs Links:
Condition 8. Every station except RSU in point B has only one RSU with
bigger coordinate value which establish a connection to:
k+1
∀i = 0 . . . m + 1, ui,j = yi , (17)
j=i+1
Condition 9. Every station except RSU in point A has only one RSU with less
coordinate value which establish a connection to:
i−1
∀i = 0 . . . m + 1, ui,j = yi , (18)
j=0
That limitations means every RSU except of RSUs in points A and B has exactly
two neighbours one - on the left and one on the right.
Condition 13.
⎧
⎪
ui,k + si,j − zi,j,k ≤ 1, ⎨∀i = 0 . . . m + 1,
∀j = 1 . . . n, (23)
2zi,j,k − ui,k − si,j ≤ 0, ⎪
⎩
∀k = 0 . . . m + 1,
− + −
Φ(ρ+
i (0), ρi (0)) = max(Φ(ρi , ρi )), (24)
for solution of linear programming problems using the full power of advanced
algorithms. The GLPK has the ability to solve problems in many different meth-
ods, such as simplex method, branch and bound method and many others. In
this case- the case of the mixed problem of linear programming (Mixed Inte-
ger Problem)-branch and bound method is used. For each ongoing experiment
GLPK-project consists of two text files: file with a description of the model
(common for all experiments) and model data file generated by Java application
from user input. Both files are written in the programming language MathProg
(GNU MathProg Language)-main language for modeling of linear programming
solvers (for example, glpsol).
Results of the problem solution by Branch and bound method are presented
in Table 1. Time characteristics expressed in seconds. Dashes means that too
much time took calculation of problem with appropriate dimension (reasonable
time is 10 min).
GLPK successfully copes with Problem 8 of dimensions 0 < |P | < 50,
0 < |T | < 5. It needs to note, that parameter |T | has more influence to calcula-
tion time than parameter |P |. Heuristic algorithm based on gradient descent is
developed for solving bigger dimension problems.
5 Conclusion
References
1. Vishnevsky, V.M., Portnoi, S.L., Shakhnovich, I.V.: WiMAX Encyclopaedia. Way
to 4G, p. 470. Tekhnosfera, Moscow (2010)
2. Vishnevsky, V.M., Semenova, O.V.: Polling Systems: Theory and Applications for
Broadband Wireless Networks, p. 317. Academic Publishing, London (2012)
Optimization of Topological Structure 39
1 Introduction
Information sharing in a smart space employs the subscription operation for
detection of content changes and for subsequent delivery of notifications to
clients [10,13]. Both change detection and notification delivery are subject to
losses in networked environments [14], especially with the emergency of Internet
of Things (IoT).
In existing solutions for smart spaces, the major role is played by information
brokers [5,9]. In this paper, we study a solution closer to the Internet philosophy,
in which the control of notification delivery is partially delegated to clients. The
latter actively request the broker for new notifications, in addition to default
passive waiting for incoming notifications from the broker. A key performance
parameter is the check interval, whose length is adapted to the observable loss
rate. This issue resembles the “classical” congestion control in Transmission
Control Protocol (TCP) [1–3], by which the window size is reduced in case
of losses and incremented otherwise.
Our study continues research [8,14]. We consider an active control of notifi-
cation delivery for subscription operation for the case of notification losses. The
problem was initially stated in [14]. In our solution, the client follows an adap-
tive strategy controlling the check interval based on the number of notifications
lost in the latest window. This adaptive strategy was early experimented in [8].
In this extended paper we show that the strategy is a generalization of the
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 40–51, 2016.
DOI: 10.1007/978-3-319-30843-2 5
Control Strategies of Subscription Notification Delivery in Smart Spaces 41
operability faults. SIB does not check delivery for already sent notifications, and
a new notification can be sent although the underlying network connection is
broken on the client side.
The above properties do not ensure the dependable notification delivery even
if reliable network protocols are used, such as TCP. A possible solution is for
a client to have an additional mechanism reducing the number of undelivered
notifications. The obvious way is augmenting the passive notification delivery
with an active control strategy that the client performs individually on its own.
Consider the following model to formalize the key properties of the subscrip-
tion notification loss problem in smart spaces. Let i = 1, 2, . . . be the event-based
time evolution on the client side, where i is the index of notification events.
An event i is either a passive notification (i.e., received from SIB) or an explicit
check of the notification delivery (made by the client within its active control).
Denote by ti and ki the time elapsed and the number of losses occurred
between i and i + 1, respectively. Assume that some initial value t0 is always
defined. The values for ki are non-negative integers. We consider the following
distributions to model the notification losses.
1. Let the time elapsed between consecutive losses follow a uniform distribu-
tion U { }. Hence, the average number of losses in any check interval is
proportional to its length ti
2. Let ki follow a Poisson process of parameter λti . Hence, the number of losses
during ti has the probability mass function
(λti )k −λti
P(ki = k) = e (1)
k!
3. Let ki follow a two-state alternated Poisson process: the loss rate is λ1 and
λ2 with probability p1 and p2 = 1 − p1 , respectively. The assumptions 0 <
p2 < p1 and λ1 λ2 describe that the network typically operates with
moderate losses (λ1 ), while from time to time the network suffers from high
losses (λ2 ), e.g., due to burst overload.
3 Control Strategies
The notification loss problem is similar to the packet loss problem in TCP conges-
tion control [1–3]. The additive–increase/multiplicative–decrease (AIMD) algo-
rithm is a feedback control algorithm used in the TCP congestion avoidance
mechanism. The AIMD algorithm calculates the size ω(τ ) of congestion window
at time τ , after the initial slow start phase.
Denote by τi the end of the ith TCP round, whose duration is then given by
ti = τi − τi−1 and by ω(τ ), where the latter is expressed in MSS units (maximum
segment size). The following equation describes the process [2]:
ω(τi ) + 1 if no loss,
ωi+1 = ω(τi + 0) = ω(τi ) (2)
if losses are observed on [τi−1 , τi ],
κ
Control Strategies of Subscription Notification Delivery in Smart Spaces 43
multiplicative
ω(τ )
decrease
multiplicative
decrease
This adaptive strategy refines the AIMD algorithm of TCP, i.e., providing
a TCP-like control of notification delivery. One can interpret ti in (3) as TCP
congestion window. Fixed δ = 1 means the additive increment by one full-sized
segment. In the AIMD algorithm, α = 0. TCP congestion implies ki−1 = 1, and
κ = ki−1 + 1 = 2 halves the value of ti , as it is described in (3).
In order to evaluate and compare the performance, we consider alternative
control strategies and present some auxiliary analytical results. The simplest
approach is the strategy of a constant check interval, i.e., ti = t0 for i = 1, 2, . . ..
The mean and variance of the random variable K, which describes the number
of losses, depend only on the loss distributions introduced in Sect. 2. In the case
of Poissonian losses, we have:
E K = Var K = λt0 . (4)
q0 q1 q2 q3
p0 0 1 2 3 ···
p1
p2
p3
p4
Note that the sum converges under reasonable assumptions over the loss
processes. (It is enough to assume that the loss probability goes to zero as
n → ∞.) The state probabilities permit to find all the relevant statistics, such
as the average check interval
∞ ∞
E T = πn t n = t 0 πn 2−n .
n=0 n=0
The assumption that the number of losses does not depend on the length of
the check interval might be unrealistic. Nevertheless, it provides a lower bound
for the average (and an upper bound for the variance) w.r.t. the more realistic
assumptions described in Sect. 2.
Moreover, a practical implementation does not use very short check intervals.
This property corresponds to truncating the Markov chain at state N , remaining
in that state in case of further losses (returning to state 0 with the probability pN
as usual). Hence, (7) is still valid for all the states up to N − 1, while
N −1
1
πN = π0 qi ,
1 − qN i=0
4 Simulation Experiments
In the experiments, we use the three notification loss distributions (see Sect. 2).
The key simulation parameters are summarized in Table 1. In all four control
strategies, we take t0 = 20 s. According to the selected parameters of the loss
distributions, it means that every 20 s one notification is lost on average. The
simulation parameters of the four control strategies are presented in Table 2.
The average check interval is exactly t0 = 20 for the strategies of constant
check interval and random selection. The adaptive strategy starts from t0 and
then forgets about t0 , balancing in accordance with the observed number of
losses. The multiplicative–decrease strategy attempts to be below t0 when losses
are observed, leading to shorter check interval on average.
Experimental behavior of the four control strategies for the given three
losses distributions is shown in Figs. 3, 4, and 5. From the visually compari-
son, it is immediately clear that the adaptive strategy makes the check interval
longer compared with the other strategies. From the point of view of the per-
formance, the adaptive strategy saves the resources, reducing the number of
network requests from the client to SIB.
Consider quantitative performance indexes. The basic metrics are the average
number of losses kavg and the average length of the check interval tavg :
n n
1 1
kavg = ki , tavg = ti .
n i=1
n i=1
Intuitively, the client is interested in reducing the losses (i.e., kavg → min) for
high values of the check interval (i.e., tavg → max). Since these two requirements
are in opposition, in analogy with the concept of power in computer networks
defined as the ratio between the throughput and the delay [12], we use the ratio
among the two, such as the following metrics:
n
kavg 1 ki
Qtot = → min , Qavg = → min .
tavg n i=1
ti
5 Conclusion
This paper has studied the problem of efficient notification delivery for the sub-
scription operation in smart spaces. We considered different assumptions on the
50 D.G. Korzun et al.
References
1. Allman, M., Paxson, V., Blanton, E.: TCP Congestion Control. RFC 5681 (Draft
Standard), September 2009. http://www.ietf.org/rfc/rfc5681.txt
2. Bogoiavlenskaia, O.: Discrete model of TCP congestion control algorithm with
round dependent loss rate. In: Balandin, S., Andreev, S., Koucheryavy, Y. (eds.)
NEW2AN/ruSMART 2015. LNCS, vol. 9247, pp. 190–197. Springer, Heidelberg
(2015)
3. Callegari, C., Giordano, S., Pagano, M., Pepe, T.: A survey of congestion control
mechanisms in linux TCP. In: Vishnevsky, V., Kozyrev, D., Larionov, A. (eds.)
DCCN 2013. CCIS, vol. 279, pp. 28–42. Springer, Heidelberg (2014)
4. Eugster, P.T., Felber, P.A., Guerraoui, R., Kermarrec, A.M.: The many faces of
publish/subscribe. ACM Comput. Surv. 35, 114–131 (2003)
5. Honkola, J., Laine, H., Brown, R., Tyrkkö, O.: Smart-M3 information sharing plat-
form. In: Proceedings of the IEEE Symposium on Computers and Communications
(ISCC 2010), pp. 1041–1046. IEEE Computer Society, June 2010
6. Kleinrock, L.: Queueing Systems: Theory, vol. 1. Wiley Interscience, New York
(1975)
7. Korzun, D.: Service formalism and architectural abstractions for smart space appli-
cations. In: Proceedings of the 10th Central & Eastern European Software Engi-
neering Conference in Russia (CEE-SECR 2014). ACM, October 2014
8. Korzun, D., Pagano, M., Vdovenko, A.: A TCP-like control of notification delivery
for subscription operation in smart spaces. In: Proceedings of the 18th Interna-
tional Conference on Distributed Computer and Communication Networks: Con-
trol, Computation, Communications (DCCN-2015), pp. 10–18. ICS RAS, October
2015
9. Korzun, D.G., Kashevnik, A.M., Balandin, S.I., Smirnov, A.V.: The Smart-M3
platform: experience of smart space application development for internet of things.
In: Balandin, S., Andreev, S., Koucheryavy, Y. (eds.) NEW2AN/ruSMART 2015.
LNCS, vol. 9247, pp. 56–67. Springer, Heidelberg (2015)
10. Lomov, A.A., Korzun, D.G.: Subscription operation in Smart-M3. In: Balandin,
S., Ovchinnikov, A. (eds.) Proceedings of the 10th Conference of Open Innovations
Association FRUCT and 2nd Finnish-Russian Mobile Linux Summit, pp. 83–94.
SUAI, November 2011
Control Strategies of Subscription Notification Delivery in Smart Spaces 51
11. Morandi, F., Roffia, L., D’Elia, A., Vergari, F., Cinotti, T.S.: RedSib: a Smart-M3
semantic information broker implementation. In: Balandin, S., Ovchinnikov, A.
(eds.) Proceedings of the 12th Conference of Open Innovations Association FRUCT
and Seminar on e-Tourism, pp. 86–98. SUAI, November 2012
12. Peterson, L.L., Davie, B.S.: Computer Networks: A Systems Approach. The Mor-
gan Kaufmann Series in Networking, 5th edn. Morgan Kaufmann Publishers Inc.,
San Francisco (2012)
13. Smirnov, A., Kashevnik, A., Shilov, N., Oliver, I., Balandin, S., Boldyrev, S.:
Anonymous agent coordination in smart spaces: state-of-the-art. In: Balandin, S.,
Moltchanov, D., Koucheryavy, Y. (eds.) ruSMART 2009. LNCS, vol. 5764,
pp. 42–51. Springer, Heidelberg (2009)
14. Vdovenko, A.S., Korzun, D.G.: Active control by a mobile client of subscription
notifications in smart space. In: Proceedings of the 16th Conference Open Innova-
tions Framework Program FRUCT, pp. 123–128. IEEE, ITMO University, October
2014
15. Weiss, N., Holmes, P., Hardy, M.: A Course in Probability. Pearson Addison Wesley,
New York (2005)
Modeling and Performance Analysis
of Interconnected Servers in a Cloud Computing
System with Dynamic Load Balancing
Udo R. Krieger(B)
1 Introduction
In recent years modern Web services have been provided by cloud computing
systems that are hosted by powerful infrastructures in modern data centers like
Microsoft Windows Azure or Amazon Web Services. The latter environment
constitutes an example of an infrastructure-as-a-service (IaaS) system where
users can deploy their virtualized service computing systems on the physical
resources of the infrastructure provider. Pooling the virtualized resources offers
the chance to follow efficiently the dynamically changing demand curves of the
Web services advertised by a service provider, and to satisfy the scalability,
elasticity, and resilience requirements of service-oriented computing.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 52–60, 2016.
DOI: 10.1007/978-3-319-30843-2 6
Modeling and Performance Analysis of Interconnected Servers 53
Fig. 1. Principle of mutual overflow routing between two coupled loss systems of vir-
tualized machines that are interpreted as trunk group 1 and 2 (cf. [1]).
Due to the dynamic MOR-type load balancing subject to the server capacities
and utilization states, we assume that this offered traffic with rate λ is randomly
split up into two portions for system 1 and 2 with rates λ1 and λ2 . Then the
described policy can be modelled by the following splitting probabilities
N1 N2
N1 +N2 (1 − ρ1 (N1 )) N1 +N2 (1 − ρ2 (N2 ))
p= , 1−p= (1)
1 − ρ(N1 + N2 ) 1 − ρ(N1 + N2 )
56 U.R. Krieger
of the Poisson stream with rate λ offered to the consecutive virtualized server
groups on system 1 and 2. Here ρ1 (N1 ) = E(X1 )/N1 , ρ2 (N2 ) = E(X2 )/N2
denote the utilization of a single virtual machine on server 1 and 2, respectively,
where the state variable Xk records the number of active virtual machines on
server k ∈ {1, 2}.
N1 N2
ρ(N1 + N2 ) = E(X1 + X2 )/(N1 + N2 ) = · ρ1 (N1 ) + · ρ2 (N2 )
N1 + N 2 N1 + N 2
is the utilization of a single machine of the same type on two adjacent servers that
are coupled by mutual overflow load balancing. Here we assume that the systems
1 and 2 will not be fully utilized. The latter case may be simply incorporated by
adopting the modified splitting probability
N1 N2
N1 +N2 (1 − ρ1 (N1 ) + ε) N1 +N2 (1 − ρ2 (N2 ) + ε)
p= , 1−p=
1 − ρ(N1 + N2 ) + ε 1 − ρ(N1 + N2 ) + ε
for sufficiently small ε > 0. It yields a static splitting according to the relative
number of VMs for fully utilized service blocks in the coupled VM clusters.
Then the resulting fresh Poisson stream of type j to system 1 has the rate
N1
N1 +N2 (1 − ρ1 (N1 ))
λ1 = λ · p = λ · (2)
1 − ρ(N1 + N2 )
and that one offered to system 2 has the rate
N2
N1 +N2 (1 − ρ2 (N2 ))
λ2 = λ · (1 − p) = λ · (3)
1 − ρ(N1 + N2 )
The flow analysis of the interacting server systems coupled by the adaptive
routing which is induced by the mutual overflow load balancing now yields the
following offered traffic rates L1 and L2 on system 1 and 2, respectively:
N1
B1 = f1 (N1 , L1 , B2 , p) = L1 N1 /[N1 !( L1 i /i!)] = E(N1 , λ · [1 − (1 − p) · (1 − B2 )])
i=0
N2
B2 = f2 (N2 , L2 , B1 , p) = L2 N2 /[N2 !( L2 j /j!)] = E(N2 , λ · [1 − p · (1 − B1 )])
j=0
if we assume without loss of generality that the service rates of all classes are
uniformly given by μ = 1. In [1] it has been revealed that this approximation of
the Markov-modulated overflow streams by Poisson flows with identical overflow
rates yields a very accurate approximation of the blocking behavior.
For any fixed splitting probability p ∈ (0, 1) this overall Erlang loss model
yields a system of fixed-point equations F = (f1 ◦ f2 , f2 ◦ f1 ) : I → I on
the compact unit square I = [0, 1]2 to determine the blocking probabilities
(B1 (p), B2 (p)) ∈ [0, 1]2 by
B1 = B1 (p) = E (N1 , λ · [1 − (1 − p) · (1 − E(N2 , λ · [1 − p · (1 − B1 )]))]) (6)
B2 = B2 (p) = E (N2 , λ · [1 − p · (1 − E(N1 , λ · [1 − (1 − p) · (1 − B2 )]))]) . (7)
The existence of a fixed point B ∗ (p) = (B1∗ (p), B2∗ (p)) ∈ [0, 1]2 is guaranteed by
Brower’s fixed-point theorem. In [1] it was shown that for fixed p ∈ (0, 1) the
independent offered Poisson streams with the positive rates λ1 , λ2 in (2), (3)
determine even a unique fixed point B ∗ (p) due to the monotonicity of Erlang’s
(n)
loss formula. Then they can be computed by a simple power iteration, e.g. B1 =
[f1 ◦ f2 ]n (B1 ), B1 = E(N1 , L1 (p)). Both blocking terms B1 , B2 arising as
(0) (0)
fixed point B ∗ (p) = (B1 , B2 ) in (6), (7) are coupled by the common splitting
probability p = g(B1 , B2 ) in (1) which depends in a nonlinear manner on the
individual server utilizations
1
ρ1 (N1 ) = g1 (N1 , L1 , B1 ) = E(X1 )/N1 = · L1 · (1 − B1 )
N1
1
ρ2 (N2 ) = g2 (N2 , L2 , B2 ) = E(X2 )/N2 = · L2 · (1 − B2 )
N2
and, hence, blocking probabilities in both loss systems 1 and 2, and on the server
utilization
E(X1 + X2 ) N1 L1 N2 L2
ρ(N1 + N2 ) = = · · (1 − B1 ) + · · (1 − B2 )
(N1 + N2 ) N1 + N 2 N1 N1 + N 2 N2
in the overall system. Hence, the splitting probability p ∈ (0, 1) in (1) is deter-
mined by the resulting fixed-point equation:
N1 L1
N1 +N2 (1 − N1 (1 − B1 ))
p = h(B1 , B2 , p) =
1− [ N1N+N
1 L1
2 N1
(1 − B1 ) + N1N+N
2 L2
2 N2
(1 − B2 )]
N1 − L1 (p)(1 − B1 (p))
= (8)
N1 − L1 (p)(1 − B1 (p)) + N2 − L2 (p)(1 − B2 (p))
N1 − λ · [1 − (1 − p) · (1 − B2 (p))](1 − B1 (p))
= (9)
(N1 − λ · [1 − (1 − p) · (1 − B2 (p))](1 − B1 (p))
+N2 − λ · [1 − p · (1 − B1 (p))](1 − B2 (p)))
58 U.R. Krieger
The corresponding fixed-point model (4), (5), (6), (7) and (9) of the combined
splitting-blocking model X = (B1 (p), B2 (p), p) ∈ [0, 1]3 is simple, but analyti-
cally complex due to the ratio structure of the term p. It can be solved by a power
iteration p(n) = h(B1 , B2 , p(n−1) ), n ∈ N, p(0) = 0.5 whose local convergence to
a fixed point X ∗ = (B1∗ (p∗ ), B2∗ (p∗ ), p∗ ) is guaranteed by Brower’s fixed-point
theorem. Starting with the outcome of our previous analysis [1], we can reveal
the dependence on the splitting parameter p∗ by a
Steady-State Representation Result. We consider the basic IaaS compo-
nent of two servers that host N1 and N2 virtual machines (VMs) and serves
Poisson arrival streams with offered loads λ1 = λp∗ and λ2 = λ(1 − p∗ ), respec-
tively. They are coupled by mutual overflow routing combined with a least-load
balancing scheme with splitting probability p∗ in (1). The steady-state distrib-
ution Π = (πi,j ), πi,j = limt→∞ P(X(t) = (i, j)) of the resulting ergodic loss
model X = (X1 , X2 ) ∈ [0, N1 ] × [0, N2 ] ⊂ N2 that describes the number of active
VMs in server 1 and 2 is determined by a perturbed product-form solution
N1
πi,j = ck (p∗ )gi (ρk (p∗ ), λ1 , p∗ ) · gj (−ρk (p∗ ), λ2 , p∗ )
k=0
N2
+ dl (p∗ )gi (γl (p∗ ), λ1 , p∗ ) · gj (−γl (p∗ ), λ2 , p∗ ) (10)
l=0
Combining all these coupled loss models arising from the hierarchical binary
load balancing tree, we are able to determine by these analytic means the vector
of the server-dependent blocking probabilities Bij of class j on server i and the
overall blocking probabilities Bi,∗ , B∗,j of all servers and classes as fundamental
performance metrics of the sketched interconnected service computing system.
3 Performance Study
We have investigated the loss and utilization performance of a basic building
block of the proposed IaaS computing model. It consists of two loss systems
Modeling and Performance Analysis of Interconnected Servers 59
4 Conclusions
We have investigated the basic component of a new IaaS computing model with
dynamic load balancing that is derived from mutual overflow routing (MOR)
among two physical servers with virtualized computing resources and a state-
dependent splitting of the offered traffic based on a least-load policy. We have
first derived a fixed-point model for the MOR scheme on an underlying binary
tree of Erlang loss systems that can reflect the state-dependent load balancing
policy. Then we have developed an analysis method to compute the loss per-
formance of the service system. The outcome has been demonstrated by a case
study of a single class IaaS block illustrating the dependence relation of the
traffic splitting and the utilization vector of the system.
60 U.R. Krieger
References
1. Krieger, U.R.: Analysis of a loss system with mutual overflow. In: Proceedings of
ITC-Seminar, Peking, September 1988
2. Maguluri, S.T., Srikant, R., Ying, L.: Stochastic models of load balancing and
scheduling in cloud computing clusters. In: Proceedings IEEE INFOCOM (2012)
3. Mukhopadhyay, A., Mazumdar, R., et al.: The power of randomized routing in
heterogeneous loss systems. In: Proceedings of ITC, Ghent (2015)
4. Mitzenmacher, M.: The power of two choices in randomized load balancing. IEEE
Trans. Parallel Distrib. Syst. 12, 1094–1104 (2001)
5. Turner, S.R.E.: The effect of increasing routing choice on resource pooling. Probab.
Eng. Inf. Sci. 12, 109–124 (1998)
Formalizing Set of Multiservice Models
for Analyzing Pre-emption Mechanisms
in Wireless 3GPP Networks
1 Introduction
The GBR services are real time services, e.g. voice telephony, video telephony,
or real time gaming, for which a minimum bit rate value is specified, i.e. guar-
anteed bit rate. Nevertheless, when free resources are available in the network,
the instantaneous bit rate can exceed the minimum specified one, but can not
exceed some threshold value known as the maximum bit rate (MBR).
The non-GBR services are those for which no minimum (guaranteed) bit
rate value is specified, the instantaneous bit rate can vary depending on the cell
load. It is determined with the help of the so-called aggregate maximum bit rate
(AMBR), allowing to differentiate services by the service priority levels. So, the
bit rate cannot exceed the maximum value specified for user equipments (UE-
AMBR) or determined by network characteristics (access point name AMBR,
APN-AMBR). Examples of such services include e-mail, web browsing, or inter-
active gaming.
Various radio resource management (RRM) mechanisms are used to guaran-
tee QoS. One of these mechanisms is the radio admission control (RAC) [2, 4–10]
aimed at admitting or rejecting user requests for service taking into account the
limited frequency bands and varying QoS requirements. For this purpose, the
different priority levels (allocation and retention priority, ARP) are assigned to
multimedia services and define the pre-emption RAC algorithms within the cor-
responding RAC schemes. According to 3GPP TS 23.203: “The range of the ARP
priority level is 1 to 15 with 1 as the highest level of priority. The pre-emption
capability information defines whether a service data flow can get resources that
were already assigned to another service data flow with a lower priority level.
The pre-emption vulnerability information defines whether a service data flow
can lose the resources assigned to it in order to admit a service data flow with
higher priority level. The pre-emption capability and the pre-emption vulnera-
bility can be either set to ‘yes’ or ‘no’.” In accordance with this definition of
ARP, it is evident that in case of the lack of resources already occupied by
lower priority services a new arrived higher priority request could be accepted,
at best, through service bit rate degradation (partial pre-emption) [4, 6, 7,9, 10],
or at worst, through service interruption (full pre-emption) [4, 7,8]. Other mech-
anisms such as reservation [5, 6], threshold [5], and probabilistic management
[5, 9] could also be applied to realize the priority-service discipline.
Since the number of services LTE users are interested in varies as well as
the services themselves within this specified range varies, more than 200 RAC
schemes could be set up considering all possible pairwise influences of users on
each others due to various pre-emption algorithms. Taking into account this fact
and the annual growth of mobile traffic, it seems that the increasing need for the
development of a general model considering the overall possible priority mecha-
nisms could be transformed into the need for a set of specific models reflecting
the predefined required services and mechanisms. Nevertheless, a unique nota-
tion is needed to describe at least, firstly each service within a RAC scheme, and
secondly, the pre-emption mechanism applied to each pair of services. This is
exactly the main purpose of the paper, i.e. to propose a such notation (Sect. 2)
and to illustrate its usage with an example set of specific models of RAC schemes
(Sects. 3 and 4).
Formalizing Set of Multiservice Models 63
Note that for GBR services increasing or decreasing bit rate from a GBR
value to a MBR one does not affect the mean service time. At the same time,
changing bit rate for non-GBR services results in changing the mean transfer
time. So, two types of traffics can be singled out: streaming traffic [11,12] that
is characterized by fixed values of bit rate and time and elastic traffic [2] that
is characterized by a fixed value of file size and a variable data transfer time
[3]. Traffic can be transferred in two modes – unicast mode or multicast mode.
For the last one two disciplines have been analyzed [4]: T1 – multicast session
is closed when the first user who has opened the session leaves it, – and T2 –
multicast session is closed when the last user leaves the session. T stands for the
so-colled “transparent” service discipline in queuing theory.
64 K. Samouylov et al.
Each of fifteen services could be described with the help of the following
parameters:
Let us consider an LTE cell having the capacity of C bps. The incoming
requests for service generate unicast, multicast, and elastic traffics and arrive as
Formalizing Set of Multiservice Models 65
Let us consider the first model (Fig. 2) [4]. Suppose that users are provided
with two services that generate streaming traffic. The service that has a higher
priority level (S1 ) (e.g. video conference) generates multicast traffic, and the
service that has a lower priority level (S2 ) (e.g. video on demand) generates
unicast traffic. The number of resources allocated for the establishment of a
multicast connection is adaptively changed along a certain specified value set of
b1 > ... > bk > ... > bK bps and is determined in accordance with the number
of free resources – max {bk , k = 1, ..., K : bk ≤ C − n}.
P 1 2
1 (0, 0) (0, 1)
2 (1, 0) (0, 0)
The admission control is realized in two stages. The first stage is the degra-
dation of the number of resources occupied by multicast traffic down to the
minimum value of bK bps. The second stage is the interruption of bK − (C − dn)
66 K. Samouylov et al.
unicast users in case of the lack of resources. The state of a multicast connection
is described as mk ∈ {0, 1} , k = 1, ..., K: mk = 1 if connection is established
and occupies bk bps, mk = 0 if connection does not occupy bk bps. The system
state space of the corresponding Markov model has the following form
X1 = {(m, n) : m = 0, n = 0, . . . , C, m = e1 , n = 0, . . . , C − b1 ,
m = ek , n = C − bk−1 + 1, . . . , C − bk , k = 2, . . . , K}.
P 1 2 3
1 (0, 0) (0, 0) (0, 1)
2 (0, 0) (0, 0) (0, 1)
3 (0, 0) (0, 0) (0, 0)
Let us denote as l ∈ {0, 1} the state of a multicast connection with discipline
T2, as b1 and b2 bps the number of resources necessary for the establishment
Formalizing Set of Multiservice Models 67
of multicast T1 and T2 connections, respectively. Then the state space has the
following form
The number of resources occupied by files can dynamically vary from a max-
imum to a minimum value that is necessary to guarantee requirements for the
mean transfer time and is realized through a threshold U of the number of trans-
ferred files. In case of low cell load, the maximum number or resources is occupied
for files transfer. If the load rises so that it is no longer possible to guarantee
a MBR for each transferred file, the instantaneous bit rate is degraded propor-
tionally to an individual MBR, in accordance with some coefficient of bit rate
degradation. Because of this, the RAC scheme and corresponding state space of
the Markov model have the following forms:
P 1 2
1 (0, 0) (1, 0)
2 (1, 0) (0, 0)
X3 = {u ≥ 0 : u1 + u2 ≤ U } .
68 K. Samouylov et al.
Fig. 6. Mean coefficient of the bit rate degradation for elastic traffic.
70 K. Samouylov et al.
To analyze the RAC scheme for the model with elastic traffic, let us define
maximum bit rate values e1 and e2 as the minimum rate values based on “group
of devices” and “technology” criterion (e.g. 4G and 3G, respectively). Note that
the ITU-T G.1010 contains information on the preferred (15 s) and acceptable
(60 s) data transfer time values for a 10 MB file, which allows us to define a limit
on the minimum elastic traffic bit rate and, consequently – on the maximum
number of the transferred data blocks U . Suppose, in this case the ratio α is
between the volume of traffic generated within a 4G network and the total volume
of traffic generated within the 2G and 3G networks. Then, for the annual range
of 2015-2018 [13], according to the Cisco Systems data, the values of α can be
obtained. Using the obtained initial data, let us conduct the analysis of the mean
coefficient of the bit rate degradation. The Fig. 6 shows that the growth of the
total offered load results in decreasing the mean degradation factor.
5 Conclusion
The paper discusses the principle of RAC schemes construction within multi-
service networks with the pre-emption based admission control. A set of three
Markov models of RAC schemes with unicast streaming, multicast streaming,
and elastic traffics was proposed. Different scenarios of priority admission con-
trol were analyzed, that are based, first and foremost, on the mechanism of bit
rate degradation and service interruption of users with lower priority levels. In
addition to the results described in the paper, a unique input data for numerical
experiments is developed for further performance analysis.
Currently, the understanding and specifications of the fifth-generation
networks are formed. In these specifications there are new types of services,
generating traffic, which does not always fit into the existing classification of
LTE network services, for example, machine-to-machine traffic. Thereby, it is
assumed that in future releases of 3GPP specifications the number of services
will be expanded again, and the formal description of RAC schemes suggested
in the paper could be also applied.
References
1. Stasiak, M., Glabowski, M., Wisniewski, A., Zwierzykowski, P.: Modelling and
Dimensioning of Mobile Wireless Networks: From GSM to LTE. Willey, Chichester
(2010)
2. Samouylov, K., Gudkova, I.: Recursive computation for a multi-rate model with
elastic traffic and minimum rate guarantees. In: IEEE International Congress
on Ultra Modern Telecommunications and Control Systems ICUMT-2010, pp.
1065–1072. IEEE Press, Moscow (2010)
3. Basharin, G., Gaidamaka, Y., Samouylov, K.: Mathematical theory of teletraffic
and its application to the analysis of multiservice communication of next generation
networks. Autom. Control Comput. Sci. 47(2), 62–69 (2013)
Formalizing Set of Multiservice Models 71
4. Borodakiy, V., Gudkova, I., Samouylov, K., Markova, E.: Modelling and perfor-
mance analysis of pre-emption based radio admission control scheme for video
conferencing over LTE. In: 6th International Conference ITU Kaleidoscope: Living
in a Converged World - Impossible Without Standards?. K-LCW-2014, pp. 53–59.
ITU Press, St. Petersburg (2014)
5. Ghaderi, M., Boutaba, R.: Call admission control in mobile cellular networks: a
comprehensive survey. Wirel. Commun. Mob. Comput. 6(1), 69–93 (2006)
6. Gudkova, I.A., Samouylov, K.E.: Modelling a radio admission control scheme
for video telephony service in wireless networks. In: Andreev, S., Balandin, S.,
Koucheryavy, Y. (eds.) NEW2AN/ruSMART 2012. LNCS, vol. 7469, pp. 208–215.
Springer, Heidelberg (2012)
7. Khabazian, M., Kubbar, O., Hassanein, H.: A fairness-based pre-emption algo-
rithm for LTE-advanced: In: 10th IEEE Global Telecommunications Conference
GLOBECOM-2012, pp. 5320–5325. IEEE Press, Anaheim (2012)
8. Kwan, R., Arnott, R., Trivisonno, R., Kubota, M.: On pre-emption and congestion
control for LTE systems. In: 72nd Vehicular Technology Conference VTC2010-Fall,
pp. 6–8. IEEE Press, Ottawa (2010)
9. Liao, H., Wang, X., Chen, H.H.: Adaptive call admission control for multi-class ser-
vices in wireless networks. In: IEEE International Conference on Communications
ICC-2008, pp. 2840–2844. IEEE Press, Beijing (2008)
10. Shorgin, S.Y., Samouylov, K.E., Gudkova, I.A., Markova, E.V., Sopin, E.S.:
Approximating performance measures of radio admission control model for non
real-time services with maximum bit rates in LTE. In: AIP Conference Proceed-
ings, vol. 1648, pp. 1–4. AIP Publishing, USA (2015). doi:10.1063/1.4912508
11. Gudkova, I., Plaksina, O.: Performance measures computation for a single link
loss network with unicast and multicast traffics. In: Balandin, S., Dunaytsev, R.,
Koucheryavy, Y. (eds.) ruSMART 2010. LNCS, vol. 6294, pp. 256–265. Springer,
Heidelberg (2010)
12. Iversen, V.: Teletraffic Engineering and Network Planning. Technical University of
Denmark, Lyngby (2011)
13. Cisco Systems: Cisco visual networking index: Global Mobile Data Traffic Forecast
Update, 20132019: usage: White paper. Cisco Systems, 40 (2015)
14. Skype 2015: How much bandwidth does Skype need?. https://support.skype.com/
en/faq/FA1417/how-much-bandwidth-does-skype-need
Methods of Performance Evaluation
of Broadband Wireless Networks Along
the Long Transport Routes
1 Introduction
One of the major problems in the development of new and operation of exist-
ing transport routes (railroads, highways, gas and oil pipes) is the creation of a
modern wireless communication infrastructure based on the international IEEE
802.11–2012 standard [1]. This standard regulates the creation of high-speed
communication channels and wireless networks that operate under the control
of IEEE 802.11n and IEEE 802.11s protocols, on the basis of which one can effec-
tively implement the wireless networks along the long transport routes. These
networks can provide not only the backbone for high-speed transmission of mul-
timedia information by deploying the base stations on high-rise buildings and
towers along the transport routes, but also an operation communication between
the fixed and mobile users (cars, trains, road signs, weight control points and
transport security control points, traffic lights control points, etc.) as well [2,3].
This research was financially supported by the Ministry of Education and
Science of the Russian Federation in the framework of the applied research project
№14.613.21.0020 of 22.10.2014 (RFMEFI61314X0020).
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 72–85, 2016.
DOI: 10.1007/978-3-319-30843-2 8
Methods of Performance Evaluation of Broadband Wireless Networks 73
A broadband wireless network along the long transport routes is a set of base
stations, connected with each other with high-speed wireless communication
channels. An adequate mathematical model for such a network is a multi-stage
queueing system with a MAP input flow, PH-distribution of service time at the
stages of the system and a cross-traffic (Fig. 1).
The term “service” of each message refers to a number of technical processes
in a real network, the duration of which is a random variable. Thus the mes-
sages are processed by one or several program components and the processing
time depends on the current load of the central processing unit (CPU) and the
memory of a base station, on the number of concurrently served messages,
the number of CPU cores and other parameters. The messages then come to
the output device and are transmitted through the network over one or more
serial communication channels to the next base station. The transmission time
via communication channels is also a random variable, since it is influenced by
the background traffic, the implemented networking technologies, the setup of
traffic profiling, and many other factors.
(0) (1)
where λij — hidden transition rates, λij — observable transition rates, and
λi — total rate of leaving a state or generation of a message without changing
the state. The sum of matrices D0 and D1 is the infinitesimal generator of the
Markov chain:
D = D0 + D1
An elementary check, considering the definition of λi , shows that the sum of
all elements for each row of the generator D is equal to zero.
Stationary probabilities θ ∈ RW of the Markov process are calculated from
the balance equations and the normalization condition:
Methods of Performance Evaluation of Broadband Wireless Networks 75
θD = 0W
(1)
θ1W =1
T
where 0W = 0 0 . . . 0 ∈ RW — row vector of zeros and 1W = 1 1 . . . 1 ∈
RW — column vector of ones. Using the obtained stationary probability distrib-
ution, we can calculate the average rate of messages, generated by the MAP-flow,
as a mean value of a random variable equal to a cumulative rate of observable
transitions from the specified state:
⎡ ⎤
W
W
λ= ⎣θi λij ⎦ = θD1 1W
(1)
(2)
i=0 j=0
Since the base stations of the wireless network have buffers of limited capac-
ity, it is necessary to take into consideration in the mathematical model the
restrictions on queue lengths at servers of each stage.
As the result, a multi-stage queuing system M AP/P H/1/N →
•/P H/1/N → · · · → •/P H/1/N is described, which adequately describes the
functioning of a broadband wireless network with a linear topology.
arrive at the system MAP/PH/1/N, and let the service time has a phase-type
distribution Y = P H(S, τ̄ ), S ∈ RM ×M , τ̄ ∈ R1×M . Then the output flow of
(Z) (Z)
the system will be a MAP flow Z = M AP (D0 , D1 ), which transition rate
matrices D0,1 ∈ R(W M (N +2))×(W M (N +2)) have forms:
(Z)
⎛ ⎞
D̃0 B1 0 0 . . . 0 0 0
⎜ 0 R0 D̃1 0 . . . 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 R0 D̃1 . . . 0 0 0 ⎟
⎜ ⎟
⎜ 0 0 0 R0 . . . 0 0 0 ⎟
(Z)
D0 = ⎜⎜ ⎟
⎟
⎜... ... ... ... ... ... ... ... ⎟
⎜ 0 0 0 0 . . . R D̃ 0 ⎟
⎜ 0 1 ⎟
⎝ 0 0 0 0 . . . 0 R D̃ ⎠
0 1
0 0 0 0 . . . 0 0 RA (3)
⎛ ⎞
0 ... 0 0 0
⎜ IW ⊗ Ct ... 0 0 0 ⎟
⎜ ⎟
⎜ 0 ... 0 0 0 ⎟
=⎜ ⎟
(Z)
D1 ⎜ ...
⎜ ... ... ... ...⎟
⎟
⎝ 0 . . . IW ⊗ Ct 0 0 ⎠
0 ... 0 IW ⊗ Ct 0
76 V. Vishnevsky et al.
D̃0 = D0 ⊗ IM
D̃1 = D1 ⊗ IM
B1 = D1 ⊗ τ ⊗ 1M
R0 = D0 ⊗ IM + IW ⊗ S − IW ⊗ Ct
(4)
RA = (D0 + D1 ) ⊗ IM + IW ⊗ S − IW ⊗ Ct
⎡ ⎤
μ10
⎢ μ20 ⎥
Ct = ⎢ ⎥
⎣ . . . ⎦ ⊗ τ = (−S1M ) ⊗ τ
μM 0
(Z) (X) (Y )
D0 = IN ⊗ D0 + D0 ⊗ IM
(Z) (X) (Y )
(5)
D1 = IN ⊗ D1 + D1 ⊗ IM
Thus, the given theorem states, that a composition of two MAPs is again a
MAP.
The formulated above theorems are used to calculate the performance char-
acteristics of a multi-stage queuing system.
Further let’s consider a method for computation of steady-state probability
distribution. The stationary probabilities will allow to find the loss probability,
the message arrival rates, the marginal distributions of queue lengths and other
characteristics of the MAP/PH/1/N system.
The steady-state probability distribution θ ∈ RL for the output MAP flow is
the solution of the system of linear algebraic equations:
θD(o) = 0L
(6)
θ1L =1
Methods of Performance Evaluation of Broadband Wireless Networks 77
– the average number Navg of messages in the system, which is defined as the
expected value of a random variable n:
N
+1
Navg = ηn n (10)
n=0
arbitrary customer arriving at the tandem by all stations. Also, the probability
of losses at different stations and subsystems of a tandem are important for
performance evaluation of a tandem network and discovering and avoiding the
so called bottlenecks in the network.
The algorithm 1 below (in a form of pseudocode) allows to use the obtained
for MAP/PH/1/N formulae to calculate the steady-state loss probabilities,
arrival rates and average queue lengths for each station.
Let K denote the number of stations, Ψ0 = M AP (A0 , A1 ) – the flow of cross-
(i) (i)
taffic arriving at each station, Φi = M AP B0 , B1 – the disposal flow from i–th
(i) (i)
station, Ψi = M AP A0 , A1 – arrival flow into the i-th station, 1 ≤ i ≤ K. Let
the service at each station follow the PH-distribution Ω = P H (τ , S).
On account of Theorem 2 we can state that:
Ψ1 = Ψ0
(12)
Ψi = Φi−1 ⊗ Ψ0 , 2 ≤ i ≤ K
Due to the enormously large matrices even in case of MAPs with 3 states and
PH-distribution of service time with 2 phases, the exact solution was obtained
for the following two cases:
that corresponds to Erlang distribution with 2 phases and arrival rates equal
to 1, service rates equal to 2, queue length N = 2, number of stations K = 4.
Parameters of the MAP were obtained from the compilation of real data,
gathered on Moscow road traffic routes at different times of day.
The Fig. 2 shows the probability density function of time intervals between
vehicles, obtained from the collected data and the probability density function
of the MAP with following matrices:
⎛ ⎞ ⎛ ⎞
−0.85 0.85 0.0 0.0 0.0 0.0
D0 = ⎝ 0.0 −1.1 0.2 ⎠ , D1 = ⎝ 0.9 0.0 0.0 ⎠
0.0 0.5 −4.0 3.5 0.0 0.0
The MAP fits the statistical data on the first two moments.
For both cases Ploss (loss probability at each server) and Pbusy (probability
that a server is busy) were found using both simulation and exact analytical
calculation.
The comparison results are given in the Table 1.
It can be seen from the table that the results of analytical modelling and
simulation have close agreement.
80 V. Vishnevsky et al.
sim analytic
Table 1. Comparison of system characteristics: Ploss and Ploss – stationary loss
sim analytic
probabilities, obtained from simulation and analytical models; Pbusy and Pbusy –
stationary probabilities that the server is busy.
starting from the case of high performance hardware and network (μ = 1000) and
proceeding to the case of a very weak hardware and low-performance network
(μ = 2). The last case appears to be relevant for the reason that under condition
of poor communication channels between stations and in case of often downtimes,
the average daily rate of service degrades significantly.
Figure 3 shows the average end-to-end delay for each station. The horizon-
tal axis depicts the number of a station that performs the first transition of a
message. The vertical axis depicts the stationary average time till this message
leaves the system (when served at the last station). The figure shows that start-
ing from μ = 100 the delay doesn’t exceed 200 ms. But if μ = 20 it takes about
1.4 s to transfer a message from the most descend station.
Figure 4 shows the average message loss ratio for each station. As it can be
seen from this figure, there are almost no losses starting from μ = 20. However,
due to the random nature of traffic, it is preferable to have a higher performance
in case of a large number of consequent stations.
Figure 5 shows the stationary non-delivery ratio of messages, sent by a spec-
ified station on their way to the last station. It is evident, that this value is a
non-increasing function of a station number, and starting from μ = 20 it is equal
to zero nearly everywhere.
Figure 6 shows the average queue length for each station. For the reason that
there additionally arrives a cross-traffic flow at each station, the arrival rate
does not decrease as the station number increases. And as all servers have the
same rate of service, the average queue length turns out to be a non-decreasing
function of the station’s number. As it can be seen from the figure, the higher
is the service rate, the later the queue gets filed up. Particularly, in case μ = 20
the queue of each station remains empty most of the time.
Finally, Figs. 7 and 8 illustrate the distributions of service delays at each
station and distributions of time intervals between the outgoing messages at
82 V. Vishnevsky et al.
each station. The figures are given for different service rates and several stations.
In particular, Fig. 8 displays that starting from a certain stage, which depends
on the service rate, the output flows become almost indistinguishable.
It is evident, that at a low rate of service the system deteriorates very fast.
It is connected mostly with the presence of cross-traffic: each consequent base
station has not only to process the messages from the previous station, but the
cross-traffic messages as well.
Particularly, the results show that for the high system performance it is neces-
sary to pay attention to the quality of communication links between base stations.
For a stable work the average rate of service should be not less than 20 messages
per second. The rate of about 100 messages per second is desirable.
7 Summary
References
1. 802.11-2012 IEEE standard for information technology. Telecommunications, infor-
mation exchange between local, metropolitan area networks. Specific Requirements
Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY)
Specifications – IEEE Std., March 2012
2. Vishnevsky, V.M., Portnoi, S.L., Shakhnovich, I.V.: WiMAX encyclopaedia. Way
to 4G. Tekhnosfera, Moscow, p. 470 (2010) (in Russian)
3. Vishnevsky, V.M., Semenova, O.V.: Polling Systems: Theory and Applications for
Broadband Wireless Networks, p. 317. Academic Publishing, London (2012)
4. Wu, Q., Zheng, J.: Performance modeling and analysis of IEEE 802.11 DCF based
fair channel access for vehicle-to-roadside communication in a non-saturated state.
Springer Wirel. Netw. 21(1), 1–11 (2014)
5. Chakraborty, S., Nandi, S.: IEEE 802.11s mesh backbone for vehicular commu-
nication: fairness and throughput. IEEE Trans. Veh. Technol. 62(5), 2193–2203
(2013)
6. Campolo, C., Cozzetti, H.A., Molinaro, A., Scopigno, R.: Augmenting vehicle-to-
roadside connectivity in multi-channel vehicular ad hoc networks. J. Netw. Com-
put. Appl. 36(5), 1275–1286 (2013)
7. Vishnevsky, V.M., Larionov, A.A., Semenova, O.V.: Performance evaluation
of the high-speed wireless tandem network using centimeter and millimeter-
wave channels. Probl. Upr. (4), 50–56 (2013). http://www.mathnet.ru/links/
a66b2ecaada565991c942fd3cc5149cf/pu800.pdf
8. Vishnevsky, V., Semenova, O., Dudin, A., Klimenok, V.: Queueing model with
gated service and adaptive vacations. In: IEEE International Conference on Com-
munications Workshops, ICC Workshopps 2009, p. 5, June 2009
9. Klimenok, V., Dudin, A., Vishnevsky, V.: Tandem queueing system with correlated
input and cross-traffic. In: Kwiecień, A., Gaj, P., Stera, P. (eds.) CN 2013. CCIS,
vol. 370, pp. 416–425. Springer, Heidelberg (2013)
10. Vishnevsky, V.M., Dudin, A.N., Semenova, O.V., Klimenok, V.I.: Performance
analysis of the BMAP/G/1 queue with gated servicing and adaptive vacations.
Perform. Eval. 68(5), 446–462 (2011)
11. Klimenok, V., Dudin, A., Vishnevsky, V.: On the stationary distribution of tandem
queue consisting of a finite number of stations. In: Kwiecień, A., Gaj, P., Stera, P.
(eds.) CN 2012. CCIS, vol. 291, pp. 383–392. Springer, Heidelberg (2012)
12. Gaidamaka, Y., Zaripova, E.: Comparison of polling disciplines when analyzing
waiting time for signaling message processing at SIP-server. In: Dudin, A., et al.
(eds.) ITMM 2015. CCIS 564 (Communications in Computer and Information
Science), pp. 358–372. Springer International Publishing, Switzerland (2015)
On the Waiting Time in the Discrete
Cyclic–Waiting System of Geo/G/ 1 Type
Laszlo Lakatos(B)
1 Introduction
We shortly repeat Koba’s results [2] to find the waiting time distribution in the
cyclic-waiting system.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 86–93, 2016.
DOI: 10.1007/978-3-319-30843-2 9
On the Waiting Time in the Discrete Cyclic-Waiting System 87
Let tn denote the time of arrival of the nth customer; its service will begin
at the moment tn + T · Xn , where T is the cycle time and Xn is a nonnegetive
integer. Let ξn = tn+1 −tn and ηn the service time of nth customer. Furthermore,
let Xn = i, if
(k − 1)T < iT + ηn − ξn ≤ kT (k ≥ 1),
then Xn+1 = k, and if iT + ηn − ξn ≤ 0, then Xn+1 = 0. Hence, Xn is a
homogeneous Markov chain with transition probabilities pik , where
pik = P {(k − i − 1)T < ηn − ξn ≤ (k − i)T }
if k ≥ 1, and
pi0 = P {ηn − ξn ≤ −iT }.
Introduce the notations
fj = P {(j − 1)T < ηn − ξn ≤ jT }, (1)
−i
pik = fk−i if k ≥ 1, pi0 = fj = fˆi . (2)
j=−∞
Let us divide the cycle time T into n equal parts called slots. For each slot
a new customer may enter with probability r, there is no entry with probability
1 − r; the service time is k slots with probability qk . Denoting the interarrival
time by ξ, the service time by η, their distributions are
P {ξ = k} = (1 − r)k−1 r, P {η = k} = qk (k ≥ 1),
i.e. they have geometrical and general distributions, respectively.
Our result is formulated in the following
Theorem 1. Let us consider the above described system and introduce a Markov
chain whose states correspond to the waiting time (in the sense the waiting time
is the number of actual state multiplied by T ) at the arrival time of customers.
The matrix of transition probabilities for this chain is
⎡ 0 ⎤
⎢ j=−∞ fj f 1 f 2 f3 f4 . . . ⎥
⎢ ⎥
⎢ −1 ⎥
⎢ f f f f f . . . ⎥
⎢ j 0 1 2 3 ⎥
⎢ j=−∞ ⎥
⎢ ⎥
⎢ −2 ⎥
⎢ fj f−1 f0 f1 f2 . . . ⎥
⎢ j=−∞ ⎥
⎢ −3 ⎥
⎢ ⎥
⎢ f f f f f . . . ⎥
⎢ j=−∞ j −2 −1 0 1 ⎥
⎣ ⎦
.. .. .. .. .. . .
. . . . . .
88 L. Lakatos
its elements are defined by (1) and (2). The generating function of the ergodic
distribution is
⎡ i
⎤
n n
i−1 (mod n) 1 − (1 − r)
∞
i
⎢ q i n −(1 − r) ⎥
⎢ i=1 1 − (1 − r)n ⎥
⎢
P (z) = ⎢1 − ⎥
Q1 (1 − r)n ⎥
⎣ ⎦
(1 − r)[1 − (1 − r)n ]
Q1 Q1 [1 − (1 − r)n ] z
−
1−r 1−r z − (1 − r)n
× n , (3)
Q1 [1 − (1 − r) ] z
1 − F+ (z) −
1−r z − (1 − r)n
where
∞
∞
F+ (z) = fj z j , Q1 = qj (1 − r)j ;
j=1 j=1
P {ξ = k} = (1 − r)k−1 r, P {η = k} = qk ,
where
∞
Qi = qk (1 − r)k .
k=i
Furthermore, let
−j
∞
Q1 [1 − (1 − r)n ] Q1
fˆj = fi = (1 − r)in = (1 − r)jn .
i=−∞ i=j
1−r 1−r
By using the transition probabilities fj , the ergodic probabilities are the solution
of the system of equations
p0 = p0 fˆ0 + p1 fˆ1 + p2 fˆ2 + p3 fˆ3 + . . .
p1 = p0 f1 + p1 f0 + p2 f−1 + p3 f−2 + . . .
p2 = p0 f2 + p1 f1 + p2 f0 + p3 f−1 + . . .
..
.
Multiplying the j-th equation by z j , summing up from zero to infinity, for the
∞
generating function pj z j we have (as in [1])
j=0
∞
j−1
∞
j
P (z) = P (z)F+ (z) + pj z f−i z −i + pj fˆj , (5)
j=1 i=0 j=0
where F+ (z) is the generating function of positive jumps, because of its com-
plexity its determination is given later.
In this expression
j−1
j−1
Q1 [1 − (1 − r)n ]
f−i z −i = (1 − r)in z −i
i=0 i=0
1−r
j
(1 − r)n
1−
Q1 [1 − (1 − r)n ] z
= ,
1−r (1 − r)n
1−
z
90 L. Lakatos
j
(1 − r)n
∞
j−1
∞
1−
Q1 [1 − (1 − r)n ] z
pj z j f−i z −i = pj z j
1−r (1 − r)n
j=1 i=0 j=1 1−
z
Q1 [1 − (1 − r)n ] z
= [P (z) − P ((1 − r)n )] ,
1−r z − (1 − r)n
∞
∞
Q1 Q1
pj fˆj = pj (1 − r)jn = P ((1 − r)n ).
j=0 j=0
1−r 1−r
this sum is represented in the table (for the sake of simplicity we omit the factor
r in all elements)
Consequently,
∞
2n
f2 = qi 1 − (1 − r)i−n−1 + qi [1 − (1 − r)n ] (1 − r)i−2n−1 ,
i=n+2 i=2n+1
Its derivative at z = 1 is
n
2n
3n kn
F+ (1) = qi + 2 qi + 3 +... + k qi + . . .
i=2 i=n+2 i=2n+2 i=(k−1)n+2
∞
∞
∞
1
+ kq(k−1)n+1 − qi (1 − r)i
k=1 k=1
(1 − r)(k−1)n+1
i=(k−1)n+1
∞
kn
∞
∞
1
= k qi − qi (1 − r)i .
k=1 k=1
(1 − r)(k−1)n+1
i=(k−1)n+1 i=(k−1)n+1
or
q1 q2 (1 − r) . . . qn (1 − r)n−1 qn+1 (1 − r)n . . . q2n+1 (1 − r)2n . . .
qn+1 . . . q2n+1 (1 − r)n . . .
q2n+1 ...
etc.
From each n columns one can factor out
qi (1 − r)i−1 (mod n)
,
1−(1−r)jn
there remain the powers of (1 − r)n , their sums are of the form 1−(1−r)n , i.e.
1−(1−r)n 1−(1−r)2n
in the first n columns 1−(1−r)n , in the second n columns 1−(1−r)n , etc. The
resulting sum will be
∞
i
(mod n) 1 − (1 − r) n n
qi (1 − r)i−1 ,
i=1
1 − (1 − r)n
and
∞
kn
∞
i
(mod n) 1 − (1 − r) n n
F+ (1) = k qi − qi (1 − r)i−1 ,
i=1
1 − (1 − r)n
k=1 i=(k−1)n+1
The substitution of this value into (6) leads to the expression (3), the condition
of ergodicity (7) gives (4).
On the Waiting Time in the Discrete Cyclic-Waiting System 93
Remark 1. By using the generating function one can compute the mean value
of waiting time (measured in cycles). In our case it is equal to
2Q1 (1 − r)n
F+ (1) −
(1 − r)[1 − (1 − r)n ] 1
C = P (1) = − ,
Q1 (1 − r)n 1 − (1 − r)n
2 − F (1)
(1 − r)[1 − (1 − r)n ] +
where
∞
kn
∞
∞
2k
F+ (1) = k(k − 1) qi − qi (1 − r)i .
(1 − r)kn+1
k=2 i=(k−1)n+1 k=1 i=kn+1
Remark 2. One can check that in the case of geometrical service time distrib-
ution qi = q i−1 (1 − q) the above formulas lead to the results of [4].
References
1. Lakatos, L., Szeidl, L., Telek, M.: Introduction to Queueing Systems with Telecom-
munication Applications. Springer, USA (2013)
2. Koba, E.V.: On a GI/G/1 queueing system with repetition of requests for service
and FCFS service discipline. Dopovidi NAN Ukrainy, no. 6, pp. 101–103 (2000) (in
Russian)
3. Lakatos, L., Efrosinin, D.: Some aspects of waiting time in cyclic-waiting systems.
Commun. Comput. Inf. Sci. 356, 115–121 (2013)
4. Lakatos, L., Efrosinin, D.: A discrete time probability model for the waiting time
of optical signals. Commun. Comput. Inf. Sci. 279, 114–123 (2014)
Quickest Multidecision Abrupt Change
Detection with Some Applications
to Network Monitoring
Igor Nikiforov(B)
1 Introduction
The quickest change detection/isolation (multidecision) problem is of impor-
tance for a variety of applications. Efficient statistical decision tools are needed
for detecting and isolating abrupt changes in the properties of stochastic signals
and dynamical systems, ranging from on-line fault diagnosis in complex tech-
nical systems (like networks) to detection/classification in radar, infrared, and
sonar signal processing. The early on-line fault diagnosis (detection/isolation) in
industrial processes (SCADA systems) helps in preventing these processes from
more catastrophic failures.
The quickest multidecision detection/isolation problem is the generalization
of the quickest changepoint detection problem to the case of K − 1 post-change
hypotheses. It is necessary to detect the change in distribution as soon as possible
and indicate which hypothesis is true after a change occurs. Both the rate of false
alarms and the misidentification (misisolation) rate should be controlled by given
levels.
2 Problem Statement
Let X1 , X2 , . . . denote the series of observations, and let ν be the serial number
of the last pre-change observation. In the case of multiple hypothesis, there are
several possible post-change hypotheses Hj , j = 1, 2, . . . , K − 1. Let Pjk and Ejk
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 94–101, 2016.
DOI: 10.1007/978-3-319-30843-2 10
Quickest Multidecision Abrupt Change Detection with Some Applications 95
denote the probability measure and the expectation when ν = k and Hj is the
true post-change hypothesis, and let P∞ and E∞ = E00 denote the same when
ν = ∞, i.e., there is no change. Let (see [1] for details)
Cγ = δ = (T, d) : min min E0 inf {Tr : dr = j} ≥ γ , (1)
0≤≤K−1 1≤j=≤K−1 r≥1
where T is the stopping time, d is the final decision (the number of post-change
hypotheses) and the event {dr = j} denotes the first false alarm of the j-th
type, be the class of detection and isolation procedures for which the average
run length (ARL) to false alarm and false isolation is at least γ > 1. In the case
of detection–isolation procedures, the risk associated with the detection delay is
defined analogously to Lorden’s worst-worst-case and it is given by [1]
ESADD(δ) = max sup esssup Ejν [(T − ν)+ |Fν ] . (2)
1≤j≤K−1 0≤ν<∞
where Cγ is the class of detection and isolation procedures with the lower bound γ
on the ARL to false alarm and false isolation defined in (1).
Another minimax approach to change detection and isolation is as follows
[2,3]; unlike the definition of the class Cγ in (1), where we fixed a priori the
changepoint ν = 0 in the definition of false isolation to simplify theoretical
analysis, the false isolation rate is now expressed by the maximal probability of
false isolation supν≥0 Pν (d = j = |T > ν). As usual, we measure the level of
false alarms by the ARL to false alarm E∞ T . Hence, define the class
Cγ,β = δ = (T, d) : E∞ T ≥ γ, max max sup Pν (d = j|T > ν) ≤ β . (4)
1≤≤K−1 1≤j=≤K−1 ν≥0
Asymptotic Theory. In this paragraph we recall a lower bound for the worst
mean detection/isolation delay over the class Cγ of sequential change detec-
tion/isolation tests proposed in [1]. First, we start with a technical result on
sequential multiple hypotheses tests and then we give an asymptotic lower bound
for ESADD(δ).
to be true.
Let Ei (N ) be the average sample number (ASN) in a sequential test (N, δ)
which chooses one of the K hypotheses subject to a K ×K error matrix A = [aij ],
where aij = Pi (accepting Hj ), i, j = 0, . . . , K − 1.
Then a lower bound for the ASN Ei (N ) is given by the following formula :
⎧ −1
⎪
⎨(1 − γ̃i ) ln K−1
α − log 2
=1 l
Ei (N )≥max ,
⎪
⎩ ρi0
−1
(1 − γ̃i ) ln βji − log 2
max
1≤j=i≤K−1 ρij
Quickest Multidecision Abrupt Change Detection with Some Applications 97
for i = 1, . . . , K − 1, where
K−1
γ˜i = γi + βi, .
=1,=i
Generalized CUSUM Test. The generalized CUSUM (non recursive) test asymp-
totically attains the above mentioned lower bound [1]. Let us introduce the fol-
lowing stopping time and final decision
Ñ = min{Ñ 1 , . . . , Ñ K−1 }; d˜ = argmin{Ñ 1 , . . . , Ñ K−1 }
of the detection/isolation algorithm. The stopping time Ñ is responsible for the
detection of hypothesis H :
Ñ = inf Ñ (k), Ñ (k) = inf n ≥ k : min Skn (, j) ≥ h
k≥1 0≤j=≤K−1
n
f (Yi )
Ñ = inf n ≥ 1 : max
min Skn (, j) ≥ h , Skn (, j) = log .
1≤k≤n 0≤j=≤K−1 fj (Yi )
i=k
The generalized matrix recursive CUSUM test, which also attains the asymptotic
lower bound, has been considered in [6,7]. Let us introduce the following stopping
time and final decision
r = min{N
N 1, . . . , N
K−1 }; d˜r = argmin{N
1, . . . , N
K−1 }
+ f (Yn )
Qn (, j) = (Qn−1 (, j) + Zn (, j)) , Zn (, j) = log
fj (Yn )
98 I. Nikiforov
subject to :
Cγ,β = δ : E∞ T ≥ γ, max max sup Pν (d = j|T > ν) ≤ β .
1≤≤K−1 1≤j=≤K−1 ν≥0
Then
log γ log β −1
SADD(N ; γ, β) max , as min{γ, β −1 } → ∞,
ρ∗d ρ∗i
details in [8,9]). For the sake of simplicity, the subscript k denoting the time
is omitted now. Let xi,j be the Origin-Destination (OD) traffic demand from
node i to node j at time k. The traffic matrix X = {xi,j } is reordered in the
T
lexicographical order as a column vector X = (x(1) , . . . , x(m) ) , where m = r2
is the number of OD flows.
Let us define an n×m routing matrix A = [a,k ] where 0 ≤ a,k ≤ 1 represents
the fraction of OD flow k volume that is routed through link . This leads to the
linear model
Y = A X,
T
where Y = (y1 , . . . , yn ) is the Simple Network Management Protocol (SNMP)
measurements. Without loss of generality, the known matrix A is assumed to be
of full row rank, i.e., rank A = n.
The problem consists in detecting and isolating a significant volume anom-
aly in an OD flow xi,j by using only SNMP measurements y1 , . . . , yn . In fact,
the main problem with the SNMP measurements is that n m. To overcome
this difficulty a parsimonious linear model of non-anomalous traffic has been
developed in the following papers [10–17].
The derivation of this model includes two steps: (i) description of the ambient
traffic by using a spatial stationary model and (ii) linear approximation of the
model by using piecewise polynomial splines.
The idea of the spline model is that the non-anomalous (ambient) traffic
at each time k can be represented by using a known family of basis functions
superimposed with unknown coefficients, i.e., it is assumed that
Xk ≈ Bμk , k = 1, 2, . . . ,
Xk = Bμk + ξk (8)
Z = W Y = W ξ (+W θ).
where
ρ∗ d = inf min ρj,0 (X j , X 0 ) and ρ∗ i = inf min min ρ,j (X , X j ).
X j ,X 0 1≤j≤K−1 X ,X j 1≤≤K−1 1≤j=≤K−1
References
1. Nikiforov, I.V.: A generalized change detection problem. IEEE Trans. Inf. Theor.
41(1), 171–187 (1995)
2. Nikiforov, I.V.: A simple recursive algorithm for diagnosis of abrupt changes in
random signals. IEEE Trans. Inf. Theor. 46(7), 2740–2746 (2000)
3. Nikiforov, I.V.: A lower bound for the detection/isolation delay in a class of sequen-
tial tests. IEEE Trans. Inf. Theor. 49(11), 3037–3047 (2003)
4. Tartakovsky, A., Nikiforov, I., Basseville, M.: Sequential Analysis: Hypothesis Test-
ing and Changepoint Detection. CRC Press, Taylor & Francis Group, Boca Raton
(2015)
5. Lorden, G.: Procedures for reacting to a change in distribution. Annals Math. Stat.
42, 1897–1908 (1971)
6. Oskiper, T., Poor, H.V.: Online activity detection in a multiuser environment using
the matrix CUSUM algorithm. IEEE Trans. Inf. Theor. 48(2), 477–493 (2002)
7. Tartakovsky, A.G.: Multidecision quickest change-point detection: previous
achievements and open problems. Sequential Anal. 27, 201–231 (2008)
8. Lakhina, A. et al.: Diagnosing network-wide traffic anomalies. In: SIGCOMM
(2004)
9. Zhang, Y. et al.: Network anomography. In: IMC 2005 (2005)
10. Fillatre, L., Nikiforov, I., Vaton, S. Sequential Non-Bayesian Change Detection-
Isolation and Its Application to the Network Traffic Flows Anomaly Detection. In:
Proceedings of the 56th Session of ISI, Lisboa, 22–29 August 2007, pp. 1–4 (special
session)
11. Casas, P., Fillatre, L., Vaton, S., Nikiforov, I.: Volume anomaly detection in data
networks: an optimal detection algorithm vs. the PCA approach. In: Valadas,
R., Salvador, P. (eds.) FITraMEn 2008. LNCS, vol. 5464, pp. 96–113. Springer,
Heidelberg (2009)
12. Fillatre, L., Nikiforov, I., Vaton, S., Casas, P.: Network traffic flows anomaly detec-
tion and isolation. In: 4th edition of the International Workshop on Applied Prob-
ability, IWAP 2008, 7–10 July 2008, Compiègne, p. 1–6 (invited paper)
13. Fillatre, L., Nikiforov, I., Casas, P., Vaton, S.: Optimal volume anomaly detec-
tion in network traffic flows. In: 16th European Signal Processing Conference
(EUSIPCO 2008), Lausanne, p. 5, 25–29 August 2008
14. Casas, P., Fillatre, L., Vaton, S., Nikiforov, I.: Volume anomaly detection in data
networks: an optimal detection algorithm vs. the PCA approach. In: Valadas,
R., Salvador, P. (eds.) FITraMEn 2008. LNCS, vol. 5464, pp. 96–113. Springer,
Heidelberg (2009)
15. Casas, P., Vaton, S., Fillatre, L., Nikiforov, I.V.: Optimal volume anomaly detec-
tion and isolation in large-scale IP networks using coarse-grained measurements.
Comput. Netw. 54, 1750–1766 (2010)
16. Casas, P., Fillatre, L., Vaton, S., Nikiforov, I.: Reactive robust routing: anom-
aly localization and routing reconfiguration for dynamic networks. J. Netw. Syst.
Manage. 19(1), 58–83 (2010)
17. Fillatre, L., Nikiforov, I.: Asymptotically uniformly minimax detection and isola-
tion in network monitoring. IEEE Trans. Signal Process. 60(7), 3357–3371 (2012)
18. Ringberg, H. et al.: Sensitivy of PCA for traffic anomaly detection. In: SIGMET-
RICS (2007)
19. Fouladirad, M., Nikiforov, I.: Optimal statistical fault detection with nuisance para-
meters. Automatica 41(7), 1157–1171 (2005)
An M/M/1 Queue with n Undesired Services
and a Desired Service
1 Introduction
In the analysis of classic queueing models it is assumed that the server is com-
pletely aware of the exact service requirement of the customer (see for example
Gross et al. [4]). It is also true that the customer knows the type of service he
needs. Thus there is no conflict on the service provided to the customer. However,
there are several real life situations where the server and/customer are (is) not
knowledgeable about the service requirement. This is especially the case where
several types of services are provided. As a concrete example we have vehicles
for repair at service stations; patients consulting physicians for diagnosis and
medication. If the right service required is not identified and instead the diag-
nosis turned out to be false the result could be disastrous. A wrong diagnosis
and consequent service provided may sometimes turn out to be even fatal/may
result in the equipment getting service, rendered totally unusable. It is this type
of problem that we analyze in this note.
We start with a simple situation. More complex situations will be analyzed
in a follow up paper. All underlying distributions assumed in this paper are
exponential. Even with this simple assumption we will soon notice that the
phase type distribution [3] gets generated in a sufficiently general setting. We
have a system providing n + 1 different services. Service time requirement for
any type during any visit is exponentially distributed with parameter depending
both the current type and the type to which service proceeds immediately after
the present. An initial probability vector for admitting to the type of service
is assumed. For clarity in exposition we call n types of services ‘undesired’ and
the remaining one, (n + 1)th type as the ‘desired’. For different customers the
‘undesired’ and ‘desired’ types could be different. Further there could be more
than one type in the desired class. These considerations will be taken up in
an extension of the present work. Thus for this paper our assumption is that n
services are ‘undesired’ (with sometimes serious consequences) and the remaining
one is the desired. However, this is hard to identify. Customers arrive according
to a Poisson process of rate λ to a single server counter. At the time when
taken for service the service requirement is correctly diagnosed with probability
θ; with complementary probability (1 − θ) the identification goes wrong. As
a consequence of correct diagnosis (prob. θ), service immediately starts with
duration having exponential distribution with parameter μ. The customer leaves
the system after service. However, if initially the customer is admitted to the
group of undesired service (with probability pi , it is diagnosed as requiring type
i service requirement, i = 1, 2, . . . , n), it may stay in this class, moving from
one undesired to another undesired, until finally all turn out to be failure and
the customer turns out to be unfit for further service (desired or undesired). It
may also happen that at some stage of service in undesired class, the service
provider identifies that the customer is being served in the undesired set of
services and so immediately takes him to the desired service stage. The customer
gets required service here and leaves. But then how long is it possible to stay
in service in the undesired set of states? We assume that a random clock (RC)
with exponentially distributed duration starts ticking the moment a customer
starts getting his service in a state in the undesired class. If correct diagnosis
is made of the desired service during its sojourn in the undesired set of states
before this random clock realizes, then the customer is immediately transferred
for service in the desired state. On completion of service, assumed exponentially
distributed with parameter μ, the customer leaves the system. On the other hand
if the RC realizes before the customer’s service need is correctly diagnosed, then
104 A. Krishnamoorthy et al.
2 Mathematical Formulation
Assumptions. The assumptions leading to the formulation of the mathematical
model are
– Arrival of customers to the system is from a Poisson stream with mean arrival
rate λ(> 0).
An M/M/1 Queue with n Undesired Services and a Desired Service 105
Let N (t) be the number of customers in the system and S(t) the service
phase at time t, which is i, 1 ≤ i ≤ n if server is at the ith undesired service
and n + 1 if at desired service. Then {(N (t), S(t)) , t ≥ 0} is a CTMC with state
space
Ω = {0} ∪ (i, j), i ∈ Z + , j = 1, 2, . . . , n, n + 1
The infinitesimal generator Q of this CTMC is a LIQBD where
⎛ ⎞
B00 B01
⎜ B10 B1 B0 ⎟
⎜ ⎟
Q=⎜ B B B ⎟
⎝ 2 1 0 ⎠
.. .. ..
. . .
with, B00 = [− λ] , B01 = λ(1 − θ)p1 λ(1 − θ)p2 · · · λ(1 − θ)pn λθ B10 =
T
α α ··· α μ , B0 = λIn+1
⎛ ⎞
α(1 − θ)p1 α(1 − θ)p2 · · · α(1 − θ)pn
αθ
⎜ α(1 − θ)p1 α(1 − θ)p2 αθ ⎟
· · · α(1 − θ)pn
⎜ ⎟
⎜ ······ ······ ··· ······ ···⎟
⎜ ⎟
B2 = ⎜ . . .. .. .. .. .... ⎟
⎜ .. .. . . . . . .⎟
⎜ ⎟
⎝ α(1 − θ)p1 α(1 − θ)p2 · · · α(1 − θ)pn αθ ⎠
μ(1 − θ)p1 μ(1 − θ)p2 · · · μ(1 − θ)pn μθ
⎛ ⎞
−(μ11 + α + λ) μ12 μ13 · · · μ1n μ1(n+1)
⎜ μ −(μ + α + λ) μ · · · μ2n μ2(n+1) ⎟
⎜ 21 22 23 ⎟
⎜ . . . .. .. .. ⎟
B1 = ⎜ .. .. .. . . . ⎟
⎜ ⎟
⎝ μn1 μn2 −(μnn + α + λ) · · · μnn μn(n+1) ⎠
0 0 ... 0 −(λ + μ)
Here,
n+1
μii = μij , 1 ≤ i ≤ n.
j=1
j=i
106 A. Krishnamoorthy et al.
We proceed with the steady-state analysis of the queueing system under study.
Naturally we have to look for the condition for stability.
and
1, i = n + 1, j = 1
C (i, j) =
0, elsewhere
Define
cnj
dnj = − ;1≤j≤n
cn,(n+1)
and for 1 ≤ k ≤ (n − 1),
k
c(n−k),j + Umj
m=1
d(n−k)j = − ;1≤j≤n
k
c(n−k),(n−k+1) + Um(n−k+1)
m=1
For 1 ≤ m ≤ k,
rl−1 −1 m
Umj = c(n−k)r1 drj −1, rj+1
rl =n−k+m+2−l j=1
1≤l≤m
r0 =n+2, rm+1 =j
An M/M/1 Queue with n Undesired Services and a Desired Service 107
Setting
kj−1 m−1
Vm = d(kj −j+1)(kj+1 −j+1) d(km −m+1)1
kj =m j=1
1≤j≤m
k0 =n
we get
n
−1
π1 = 1+ Vm
m=1
and
j
π j+1 = djk π k , for 1 ≤ j ≤ n.
k=1
The LIQBD description of the model indicates that the queueing system is
stable if and only if
ΠB0 e < ΠB2 e.
This gives the stability condition as
Lemma 1. The system is stable if and only if
Proof. One can calculate left (decrease by one in the number of customers) and
right (increase by one in the number of customers) drifts for stability, left drift
rate should be larger than right drift rate. One may also proceed heuristically
to arrive at the above result.
x Q = 0, xe = 1 (2)
R2 A2 + RA1 + A0 = 0 (4)
0.32
Ws when α=2
0.3 W when α=3
s
Ws when α=4
0.28
0.26
Ws
0.24
0.22
0.2
0.18
0.16
0 0.2 0.4 0.6 0.8 1
θ
0.8
x0
0.7 x.1
x
.2
0.6 x
.3
x.4
0.5
0.4
0.3
0.2
0.1
0
2 2.5 3 3.5 4 4.5 5
λ when α=4
References
1. Madan, K.C.: An M/G/1 queue with second optional service. Queueing Syst. 34(1–
4), 37–46 (2000)
2. Medhi, J.: A single server poisson input queue with a second optional channel.
Queueing Syst. 42, 239–242 (2002). http://link.springer.com/article/10.1023%2FA%
3A1020519830116
3. Neuts, M.F.: Matrix-Geometric Solutions in Stochastic Models. Johns Hopkins Uni-
versity Press, Baltimore (1981)
4. Gross, D., et al.: Observation of Strains: Fundamentals of Queueing Theory. John
Wiley and Sons, Inc., Hoboken (2011)
Swarm of Public Unmanned Aerial
Vehicles as a Queuing Network
1 Introduction
One of the most attractive areas of the networks and communication systems
has recently been Flying Ad Hoc Networks (FANET) [1–3]. Initially used mainly
for military purposes, UAVs are currently used in civilian applications [4, 5].
By analogy with the division of terrestrial in the Ad Hoc network [6, 7] and
ubiquitous or wireless sensor networks [8, 9] in the field of Ad Hoc networks there
were flying ubiquitous sensor network FUSN [10]. Widespread public unmanned
aerial vehicles and related networking features FUSN enable to identify a new
class of public communications networks FUSN-P (Public) [11]. One of the main
features of FUSN-P is that the UAV is operated usually by nonprofessional
users, so that it requires the simplest handling of them during operation. For
this purpose, in [11] in the FUSN-P it was proposed to use the UAV flight
for the data collection from the sensor fields on a given route. Simultaneous
use of multiple UAVs leads both to creation of a swarm and to the possibility
of considering it as a swarm of the queuing network. Notable works of UAV
swarms as a part of FANET usually pursued the target of cooperation the UAV
opportunities for solving military tasks, for search of the target, etc. [12–14].
We believe that the wide spread of public unmanned aerial vehicles enables to
consider a separate UAV as a queuing system [10] and a swarm as a queuing
network.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 111–120, 2016.
DOI: 10.1007/978-3-319-30843-2 12
112 R. Kirichek et al.
If the coordinates of the terrestrial nodes are accidental, the entry system
receives the random flow of the entities. The properties of this flow are deter-
mined by the properties of the sensor field (publishing sites on the surface), the
radius of the service drones and its speed. We will make the following assump-
tions:
– the sensory field is a Poisson field;
– the UAV is believed to move in a straight line at a constant velocity v;
– the zone service is a circle with a radius R.
Define the distribution function for the incoming flow entities. For this pur-
pose we will examine the service area of the UAV at time 0 and at time t. During
t the entities (nodes) which are found in the area that is defined by a shift of the
UAV service area for time t will go in the system. According to the properties
of the Poisson field, the probability of presence of n points (nodes) in a certain
area is determined by the Poisson distribution and depends only on the field
area. The probability of presence of z entities (nodes) in the field S is
az −a
pz = e (1)
z!
Swarm of Public Unmanned Aerial Vehicles as a Queuing Network 113
S(t) = 2R · vt (3)
The flow rate, i.e. the average number of entities per unit of time is equal to
γ = p · 2R · vt (4)
The distribution of the time interval between the entities We will consider
the random variable T as the time interval between two successive events in the
stream and will find its distribution function.
Then the probability that z entities will go to the time section of the length t is
P (T ≥ t) = 1 − F (t) (6)
Considering this fact, the distribution function of the time interval between
the entities is
F (t) = 1 − e−p·2·R·vt (8)
Thus, the elementary flow will enter the system, the time intervals between
the entities, which are distributed exponentially with a mean.
1
ā = (9)
ρ · 2R · v
Taking into consideration the above mentioned facts, the flow of entities (mes-
sages), which arrives at the node of each of the UAV has the properties of a simple
entity flow. Beside the flow of messages from a particular terrestrial sensor field,
the viewed nodes receive the traffic flows from other nodes on the network.
Further we will assume that the output flow of messages from i node with
n
probability rij is an input to the node j. With probability 1 − rij the entities
j=1
will leave the node i and will be sent to the external environment, i.e., to the
gateway, Fig. 2.
114 R. Kirichek et al.
Fig. 2. Model of data delivery route between the source (s) and receiver (t).
In the general case, the service time of the messages on the route segment t
consists of two main components: the time of sending the message on channel τ
and the time-out state of the channel readinessψ, which are generally random.
Changing of the channel status is a random process that occurs under the
influence of many independent factors (events), such as the entry and stepping
out of communication range due to the random deviations from the desired
path of movement, the effect of interference from transmitters located on the
other elements of the system and others. It is expected that with a sufficiently
large number of independent events the channel readiness intervals will have the
distribution which is close to exponential distribution, therefore, the state of
waiting time readiness ψ will also have the similar distribution.
If the time distribution of the message sending through the communication
channel τ is close to an exponential one, then the assumption of exponential
distribution of service time t is quite possible.
If we strengthen the above mentioned conditions of the network by the
assumption of exponential service time of messages in the nodes, these conditions
will coincide with the conditions of the network Jackson [16].
M
λj
T = Tj , (10)
j=1
γ
Fig. 3. Model of data delivery route between the source (s) and receiver (t).
Fig. 4. The probability density of the delivery time on the route of m length = 1,2,3,4
hops.
mk
θk = Tj (14)
j=1
Fig. 5. Random placement of 100 nodes in the cube 200 × 200 × 200 m (a) and in the
sphere of equal volume (b) 14.
Fig. 6. Distribution of the lengths of the shortest paths in the network of 100 nodes
in the cube 200 × 200 × 200 m.
118 R. Kirichek et al.
Fig. 7. The distribution of the lengths of the shortest paths in the network of 100
nodes in an equal volume area.
Figure 7 shows the distribution of the lengths of the shortest paths in the net-
work with a random arrangement of 100 nodes in the area with a communication
network node radius of 50 m.
The average path length was 5.18 hop. For comparison, the same figure also
shows a Poisson distribution with a mean of 5.18. In this case, the network
connectivity was 0.94.
The connectivity probability can be defined as the probability of falling into
the sphere of a given radius of at least one node.
Out of the properties of the Poisson field, this probability is
a=V ·ρ (16)
For the simulated network, starting from (17), it is equal to 0.999. The values
of connectivity which are obtained from the simulation results are within the
Swarm of Public Unmanned Aerial Vehicles as a Queuing Network 119
error due to the finite size of the sample. It should be noted that the expression
(17) gives the probability of connectivity for the unlimited Poisson field. In this
case, the field is limited with a certain volume. In the case of restrictions, “edge
effect” takes place which considers that the probability of connectivity for the
nodes near the border is less than for the nodes that are closer to the center
of considered limiting volume of the figure. This is obvious when considering a
node which is located strictly at the boundary field.
The adjacent to it node can be located within the area. If the boundary
is the plane, the extent to which communication with the neighboring node is
possible is less than half for the site located near the center of the examined
area (if the communication range is smaller than the area of the node). In this
regard, it should be expected that the assessment of the connection probability
(17) is the upper bound. Also the closer to the probability the value of the
connected network will be (17), the larger the ratio of bounding shape to its
surface area is. It is obvious that by increasing of the geometric dimensions the
ratio will increase. As it is seen from the given figures in the case of considering
the limited space of a cube, the length of the shortest path is well described
by a Poisson distribution. In the case when the space is limited by a sphere,
the distribution of the lengths of the shortest paths differs from the Poisson
distribution to a greater extent. The average lengths of the shortest path (in the
race) in the cases of cube and sphere are expected to vary.
4 Conclusion
1. While the organization of interaction with UAVs USN nodes to collect data
under the certain conditions, the network connections between the UAV can
be seen as a queuing network.
2. When a sufficiently large number of nodes which are located on the UAV
model, the network Jackson can be used. In this case, the delivery time of
data between the sources and receiver will obey the law of Erlang.
3. With a relatively small number of UAV to estimate the time of the data deliv-
ery it is possible to use familiar approximate estimates for systems G/G/1.
4. The number of “hops” in the shortest route between the nodes of the UAV is
distributed according to the law which is close to the Poisson law that enables
to estimate the length of the routes and the delay of the data delivery.
References
1. Bekmezci, I., Sahingoz, O.K., Temel, S.: Flying Ad-Hoc networks: a survey. Ad
Hoc Netw. 11(3), 1254–1270 (2013)
120 R. Kirichek et al.
1 Introduction
µ
... ...
µ
R H1 L1 µ
H 2 L2
H3 L3 ...
µ
HK 1 L4
In the empty system, there is only one active server. Server activa-
tion/deactivation procedures are governed by number of customers in the system
upper threshold vector H = (H1 , H2 , ..., HK−1 ), H1 < H2 < ... < HK−1 ) and
lower threshold vector L = (L1 , L2 , ..., LK−1 ), L1 < L2 < ... < LK−1 ) where
Lk+1 < Hk , k = 1, K − 2 and Lk < Hk , k = 1, K − 1. Customers are served in
FCFS (First Come First Served) order. System functions as follows:
Probability Measures of Cloud Computing Systems with Dynamic Scaling 123
– ak,i,n is probability that starting from state (k, i, n) the system will pass to
(k, i, n − 1) earlier, than to states on levels k + 1, i or k, i + 1,
– a∗k,i is probability that starting from state (k, i, Lk ) the system will pass to
(k, i, Lk − 1) earlier, than to state (k, i + 1, Lk − 1),
– Λk,i,n is intensity of direct transition from states of level k, i − 1 to a state
(k, i, n) considering that states (k, i, n + 1), (k, i, n + 2), ... are eliminated,
– Ak,i,n is probability that starting from state (k, i, Hk−1 + 1) the system will
pass to (k, i, n) earlier, than to states of levels k + 1, i or k, i + 1,
– A∗k,i,n is probability that starting from state (k, i, Hk−1 + 1) the system will
pass to (k, i, Lk − 1) earlier, than to state k, i + 1, Lk − 1,
– Ck,i is probability that starting from a state (k + 1, i, Hk + 1) the system will
pass to (k, i, Lk − 1) earlier, than to states of levels k, i + 1.
For these auxiliary probabilities and intensities recurrent relations (2–8), (11–
24), (26–32), (34–38) hold true.
Stationary probabilities for states (k, i, n) for k = 1, K − 1, i = k are calcu-
lated by formula (1).
Λk,i,n + λπk,i,n−1
πk,i,n = , Lk−1 ≤ n ≤ Hk , (1)
iμ + λ(1 − ak,i,n+1 )
πk,i,Lk−1 −1 := 1 for k = 1,
πk,i,Lk−1 −1 := 0 for k = 2, K − 1.
The auxiliary probabilities and intensities used in (1) are calculated using
recurrent formulas (2–8).
ak,i,Hk +1 := 0; (2)
iμ
ak,i,n = , Lk + 1 ≤ n ≤ Hk ; (3)
iμ + λ(1 − ak,i,n+1 )
ak,i,n+1 := 1, Lk−1 ≤ n ≤ Lk − 1, (4)
Λk,i,Hk +1 := 0; (5)
Λk,1,n := 0, Lk−1 ≤ n ≤ Hk ; (6)
Λk,i,n = απk,i−1,n + ak,i,n+1 Λk,i,n+1 , Lk ≤ n ≤ Hk , (7)
Hk
Λk,i,n = α πk,i−1,v + λ(1 − Ck,i−1 )πk,i−1,Hk , Lk−1 ≤ n ≤ Lk − 1, (8)
v=n
Probability Measures of Cloud Computing Systems with Dynamic Scaling 125
ak,i,Hk +1 := 0; (11)
iμ
ak,i,n = , Lk−1 + 1 ≤ n ≤ Lk − 2, Lk ≤ n ≤ Hk ; (12)
iμ + α + λ(1 − ak,i,n+1 )
iμ
ak,i,Lk −1 = ; (13)
iμ + α + λ(1 − a∗k,i )
λCk,i
a∗k,i = ak,i,Lk + (1 − ak,i,Lk ) , (14)
λ+α
Λk,i,Hk +1 := 0; (15)
Λk,1,n := 0, Lk−1 ≤ n ≤ Hk ; (16)
Λk,i,n = 2απk,i−1,n + ak,i,n+1 Λk,i,n+1 , Lk ≤ n ≤ Hk ; (17)
Hk
Λk,i,n = 2α πk,i−1,v + λ(1 − Ck,i−1 )πk,i−1,Hk , Lk−1 ≤ n ≤ Lk − 1 (18)
v=n
Hk +1
iμ
Ck,i = ak+1,i,v , (19)
iμ + α + λ(1 − ak+1,i,Lk +1 )
v=Lk +1
Ak,i,Hk−1 +1 := 1; (20)
Ak,i,n := 0, Hk−1 + 2 ≤ n ≤ Hk ; (21)
Hk−1 +1
Ak,i,n = ak,i,v , Lk − 1 ≤ n ≤ Hk−1 ; (22)
v=n+1
λCk,i
A∗k,i = Ak,i,Lk −1 + (1 − Ak,i,Lk −1 ) ; (23)
λ+α
Lk −1
Ak,i,n = A∗k,i ak,i,v , Lk−1 ≤ n ≤ Lk − 2. (24)
v=n+1
126 Y. Gaidamaka et al.
N= nπk,i,n . (39)
(k,i,n)∈S
Blocking probability:
K
π= πK,i,R . (42)
i=1
5 Numerical Analysis
In this section we perform a comparative analysis of initial system, proposed in
[10] and simplified model, described above. We compare two main performance
characteristics for both models, i.e. mean response time and blocking probability.
We provide calculations for the system with K = 20 servers, serving rate μ = 1
and two different arrival rate intensities λ = {12, 18} for various values of server
switching rate α, which is key parameter that makes behavior of the systems
differ from each other.
128 Y. Gaidamaka et al.
Figures 3 and 4 show mean response time of both models for thresholds
arrangements (44) and (45) respectively. As it is seen of Fig. 3, simplified and
initial model have almost the same response time for α μ
≈ 1. With the growth of
μ
α ratio to 10, relative error of the simplified model increases to approximately
2 % for low load and 1 % for high load case and reaches 10 % and 5 % when
μ
α = 100 for light and heavy loads respectively. Note that simplified model show
lower mean response time because with low additional server activation rate we
assumed activation to complete immediately in some cases.
With wider inter-threshold range arrangement (45), relative error of the sim-
μ
plified significantly decreases (Fig. 4). It becomes nearly 0.2 % in α = 10 case
μ
and 3 % in α = 100 case.
Figures 5 and 6 show blocking probability of both models for thresholds
arrangement 44 and 45 respectively. Figures show that in light load case, block-
ing probability relative error of the simplified model is extremely high even for
μ
α = 5, but it may be considered feasible in heavy load case. The reason of that
is more frequent immediate server activation, which increases system utiliza-
tion. However, in heavy load cases system shows almost maximum performance
already, and additional factors do not have significant effect.
Probability Measures of Cloud Computing Systems with Dynamic Scaling 129
6 Conclusion
In this paper, we proposed the simplified model for performance analysis of cloud
computing system with dynamic server activation. For the considered model, we
developed an efficient stationary probabilities computing algorithm based on
state elimination technique. Simplificaiton of the model significantly decreases
state space complexity of corresponding Markov process and, consequently, com-
puting complexity of the algorithm.
130 Y. Gaidamaka et al.
Our analysis showed that the simplified system gives good precision of the
mean response time for a variety of system and load parameters. Moreover,
increasing distance between corresponding lower and upper thresholds, improves
the degree of calculation accuracy. However, blocking probability precision can
be considered feasible only in heavy load cases.
Another way to increase computation accuracy of the simplified model is to
increase maximum simultaneously running server activation procedures to three
or four and it will be done in our further research.
Acknowledgments. This work was supported in part by the Russian Foundation for
Basic Research, projects No. 14-07-00090, 15-07-03051, 15-07-03608.
References
1. ETSI Cloud Standards Coordination. Final Report 2013, ver. 1.0. http://
www.etsi.org/images/files/Events/2013/2013 CSC Delivery WS/CSC-Final
report-013-CSC Final report v1 0 PDF format-.PDF
2. Goswami, V., Patra, S.S., Mund, G.B.: Performance analysis of cloud with queue-
dependent virtual machines. In: 1st International Conference on Recent Advances
in Information Technology, Dhanbad, India, pp. 357–362. IEEE Press (2012)
3. Pechinkin, A., Gaidamaka, Y., Sopin, E., Talanova, M.: Performance analysis of
cloud computing systems with dynamic scaling. In: IX International Annual Sci-
entific and Practical Conference “Modern Information Technologies and IT Edu-
cation”, pp. 395–406. INTUIT.RU, Moscow (2014)
Probability Measures of Cloud Computing Systems with Dynamic Scaling 131
Lviv Polytechnic National University, Bandera street 28a, Lviv 79015, Ukraine
ivanna.droniuk@gmail.com, mar.nazarkevych@gmail.com,
olha.fedevych@gmail.com
http://www.lp.edu.ua/ikni
1 Introduction
When it comes to protecting data in computer networks, there is often a need to
generate some noise signals. This is especially true for wireless communication
systems, including networks which use Code Division Multiple Access (CDMA)
technology [1]. CDMA technology is based on a SST (DH-SS Direct Sequence
Spread Spectrum) transfer technology, when information is as if smeared on a
wide range of frequencies. Harmonic oscillations often play the role of informa-
tion carrier. In book [2] the author describes a classical approche to describe
additive, white and Gaussian models of noise. The noise is present in all com-
munication systems and it was an obstacle for the detection of signals and has to
be eliminated by the filters. With the further development of telecommunication
technologies, the information transmission methods based on noise signals were
emerged [3]. During the building of modern confidential communication systems,
the noise signals transformed from the obstacles to the main signal carriers.
However, during the last decade, people have begun to use noise signals as the
carrier of main information transmitted by connection channels. Noise signals
have several advantages in terms of information security both in communication
cable channels and in radio channels. Data transferring in noise signal can be a
good and effective alternative for the methods of cryptographic protection.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 132–140, 2016.
DOI: 10.1007/978-3-319-30843-2 14
Synthesis of Noise-Like Signal Based on Ateb-Functions 133
It is known [1], that the carrier oscillation, formed on the basis of noise signal,
enables effective recognition of signals by form. Noise signals also has another
advantage - it gives a possibility to provide transmitting protection on the level
of physical channel, which is especially important during the construction of
protected multi-access communication systems. Noise-like signals is a type of
noise signals, where the role of carrier can be played by signal constructions
implemented on the basis of harmonic oscillations.
Oscillations that arise in nonlinear systems with multiple degrees of freedom
generalize harmonic oscillations and are described by the relevant differential
equations. Mathematically, these oscillations can be described by using Ateb-
functions. This is the idea behind this paper: to apply the method of construc-
tion of noise-like signal based on harmonic oscillations to generalize oscillations,
described by Ateb-functions, and to build noise-like oscillations based on Ateb-
functions. As noise-like signals are used in CDMA technology, among others,
the proposed approach is relevant and can have a wide practical use to protect
the transmission of information in computer networks. The proposed method
can be used for noise signal synthesis in both hardware and software domains.
The purpose of this research is to develop mathematical model based on Ateb-
functions for the problem of synthesis of discrete noise signal with given spectral
and autocorrelation properties.
Let Γ (x) denotes the Gamma function. The period Π(m, n) of Ateb-functions
shall be calculated on the following formula
Γ n+11
Γ m+11
Π(m, n) = (2)
Γ n+11 1
+ m+1 .
superellipse as [5]
x p y q
+ = 1, where p, q > 0. (3)
a b
For simplicity, let us to assume a = b = 1, and p, q are related to the para-
meters of Ateb-functions as follow p = m + 1, q = n + 1. The substitution
x = ca(n, m, ω), y = sa(m, n, ω) into Eq. (3) of a generalized superellipse, trans-
forms (3) to the identity (1). Thus it shows that the main Ateb-functions identity
(1) is the representation of a superellipse Eq. (3).
Now it is turning to the description of the method of forming a noise signal
based on the Ateb-functions. Taking into consideration characteristic features of
real noise signal, the real noise signal s(t) - is a set of simultaneously existing
power oscillations of frequency, phase, and amplitude, which are random. The
spectrum of noise signals covers a wide band of frequencies. If this spectrum
is uniform at all frequencies from 0 to ∞, then such noise is called white. In
practice, this noise cannot be obtained, but for any hardware, whose bandwidth
is in many times less than noise signal spectrum, the noise can be considered
as white. The power of noise signal used is determined by the bandwidth of the
device on the input of which it comes. If a controlled noise signal with duration
T is radiated, and acceptance is going through a concerted filter or a correlation
circuit, then during the output of correlator with the coincidence of the time of
received and reference signals (τ = 0) the energy would be released
T
E= s(t)2 dt. (4)
0
In conditions that this energy is also random variable with average mean
E0 and dispersion σ 2 , fluctuation of energy of carrier oscillation can be seen as
interference, which is distorting the signal. If the noise is limited by Δf stripe
and has a large base B, then the relative standard deviation of value is equal [6]
σ2 1
[h] ≈ . (5)
E0 Δf T
In order to minimize distortion of information transferred by noise, it is
appropriate to use signal with a large base. Then, based on (4), the results of
measurements of noise energy will become average during the time which in sev-
eral times exceeds the correlation time τ of the investigated signal. White noise
is an ideal noise signal that has an infinite spectrum and correlation function in
the form of a delta function. In practice, white noise cannot be obtained, because
equipment with unlimited pass band is needed for its generation and process-
ing. For this reason in real systems noise with limited stripe Δf is used. Work
[6] considers the method of construction of noise-like signals based on harmonic
oscillations. As Ateb-functions can generate a wider range of fluctuations that
are generalizations of harmonious and are periodic, the present work general-
izes the proposed in [4] method based on Ateb-functions. The main advantage
Synthesis of Noise-Like Signal Based on Ateb-Functions 135
Fig. 1. Dependence of period Π(m, n) from the parameters n and m, which are belong-
ing to the interval [−5; 5].
Fig. 2. Algorythm of forming a discrete noise sygnal based on sa( 17 , 17 , t) (on the left
hand superellipse is presented, on the right hand - the graph of Ateb-sine), dashed lines
correspond to discrete signal points.
dsa(n, m, t) 2
= cam (m, n, t) (7)
dt n+1
Ateb-sine function figure looks similar to cosine and is presented on Fig. 2.
Whereas |ca(•)| ≤ 1, function sa(m, n, t) grows more slowly with the increase
of parameters n and m. This property is also demonstrated on Fig. 2. With the
help of parameters n and m, the abovementioned properties allow influence the
spectrum of noise-similar signal u(m, n, t) generated on the Eq. (6).
Method of noise signal formation based on Ateb-sinus is shown on Fig. 2.
Method will be applied for forming continuous or discrete signals.This figure
shows a system of rectangular coordinates with the starting point K with coor-
dinate (1; 0) when initial phase in (6) φ(t) = 0, which moves counterclockwise
with a constant speed by superellipse (point K1 , K2 , K3 for discrete case), and
noise signal u(m, n, t) is formed by the formula (6). Graphs of synthesized signals
with differential base B are shown on Figs. 3, 4. These and next figures are cre-
ated using MatLab 7.0 computation system. In order to improve noise capacity
and to investigate the statistical characteristics of the generated signals, nor-
malized autocorrelation functions and spectrums were discussed. White noise is
a random signal with a constant power spectral density. A special type is the
Gaussian white noise (GWN), which has a Gaussian distribution amplitude. It is
a mathematical approach for describing a real noise signals. According to these
mathematical descriptions, generated by (6) broadband noise-like signals belong
to the Gaussian white noise, which sometimes is named as noise-like signals.
Statistical characteristics GWN: a mean is equal to zero, and variance is a finite.
Synthesis of Noise-Like Signal Based on Ateb-Functions 137
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 50 100 150 200
Fig. 3. The noise, generated on the basis of the Eq. (6), signal base B = 200 and Ateb
parameters n = 17 , m = 17 .
0.8
0.6
0.4
0.2
−0.2
−0.4
−0.6
−0.8
−1
0 200 400 600 800 1000
Fig. 4. The noise, generated on the basis of the Eq. (6), signal base B = 1000 and
1 1
Ateb parameters n = 20 , m = 20 .
138 I. Dronyuk et al.
100
80
60
40
20
−20
−40
0 50 100 150 200
600
500
400
300
200
100
−100
0 200 400 600 800 1000
1 1
Fig. 6. ACF of the generated noise signal B = 1000, n = 20
,m = 20
.
5 Conclusion
The mathematical model of the noise signal based on Ateb-functions was pro-
posed. The connection between Ateb-functions and the curve of superellipse was
shown. The implementation of the method of synthesizing continuous or dis-
crete signal structures with given spectral and autocorrelation characteristics
based on Ateb-functions was presented. The required properties of formed noise
implementation are provided by varying the parameters of Ateb-functions. The
application of noise signals for the purposes of information security in modern
communication systems, including the use of CDMA technology was discussed.
In future investigation it will be interesting to change amplitude, frequency
and phase of signal are changed in accordance with the distribution law of
random or pseudorandom numbers. It will be also interesting to investigate a
comparative analysis of implementations of noise signals based on different dis-
tribution laws.
140 I. Dronyuk et al.
References
1. Ipatov, V.P.: Spread Spectrum and CDMA: Principles and Applications. John Wiley
and Sons, Ltd. (2005). ISBN: 0-470-09178-9
2. Sklar, B.: Digital Communications: Fundamentals and Applications, 2nd edn. Pren-
tice Hall P T R, New Jersey (2001)
3. Rizzo, C., Brookson, C.: Security for ICT - the Work of ETSI, ETSI White Paper
No. 1, Fourth edition, January 2012
4. Grytsyk, V.V., Dronyuk, I.M., Nazarkevych, M.A.: Information technologies of doc-
ument protection by Ateb-functions means. Part 1: Building an Ateb-functions data-
base to protect documents/Problems of control and informatics, 2, 139–152 (2009).
(In Russian)
5. Sokolov, D.D.: Lame curve. In: Hazewinkel, M. (ed.) Encyclopedia of Mathematics.
Springer (2001). ISBN 978-1-55608-010-4
6. Korchinskiy, V.V.: Method of modeling of noise signals for systems of transmission
confidential information/Vestnik NTU “HPI” - Kharkiv, 38(1011), 99–104 (2013).
(In Russian)
7. Senyk, P.M.: About Ateb-functions/Rep. USSR Academy of Sciences, Ser. A, No.1,
pp. 23–27 (1968). (In Ukrainian)
Approach to Estimation of Performance
Measures for SIP Server Model
with Batch Arrivals
1 Introduction
This work was supported in part by the Russian Foundation for Basic Research,
projects No. 15-07-03051, 15-07-03608.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 141–150, 2016.
DOI: 10.1007/978-3-319-30843-2 15
142 Y. Orlov et al.
N b1 v2 l(2) − l(1) b1 b2 ρ
τ= = + + . (4)
ρ 2v1 2l(1) 1 − ρ 2b1 1 − ρ
144 Y. Orlov et al.
You can see the average queue length (Fig. 1) and average waiting time
(Fig. 2) for 4 batch size distributions for exponential service time and vaca-
tion time with the following initial data b1 = 15 ms – average service time,
v1 = 15l1 ms – average vacation time, l1 = 3 – average batch size. This is the
starting point of the reseaches. The plots (Figs. 1 and 2) have similar behavior.
We investigate the sensitivity of the model parameters to probability variation
of batch size.
As a base of our analysis we use 4 batch size distributions with following dis-
tribution functions (Table 1), where fi (k) is the probability, that batch size is
equal k. We assume, that maximum batch size is equal to 8. The distribution
n
functions can be found by formula Fi (n) = k=1 fi (k).
We consider 5 types of norms and define the distance between two DFs by the
formulas from the Table 2, where the distance from i to j distribution functions
is ρnorm
ij . The first step of our investigation was to compare distances between
DFs in several norms.
Table 3 shows the distances between batch size DFs in considered norms. For
example, the first row shows the distance between Logarithmic and Geometric
batch size distributions, the last row – the distance between Zipf and Uniform
distributions. Statistics show that the distances are dependent from each other.
So, for the next step of our investigation we can use any appropriate norm.
1 (0,85)
k 8 (0,85)k
Logarithmic DF f1 (k) = Z1 k
, Z1 = k=1 k
8
Geometric DF f2 (k) = Z12 (0, 67)k , Z2 = k=1 (0, 67)
k
8
Zipf DF f3 (k) = Z13 k , Z3 = k=1 1/k
Uniform DF f4 (k) = 1/8
C norm ρC
ij = max |fi (k) − fj (k)|
k
8
L1 norm ρL1
ij = k=1 |fi (k) − fj (k)|
2
Hellinger distance ρij = 8k=1
He
fi (k) − fj (k)
8
Kullback-Leibler distance ρKL
ij = k=1 fi (k) ln (fi (k)/fj (k))
8
Supplement for total area S ρS
ij = 1 − k=1 min (fi (k), fj (k))
Comparison C L1 He KL S
Logarithmic, Geometric 0,13 0,27 0,02 0,04 0,134
Logarithmic, Zipf 0,11 0,26 0,03 0,05 0,129
Logarithmic, Uniform 0,35 0,86 0,26 0,52 0,431
Geometric, Zipf 0,05 0,18 0,01 0,03 0,090
Geometric, Uniform 0,22 0,71 0,18 0,33 0,354
Zipf, Uniform 0,24 0,60 0,13 0,26 0,302
146 Y. Orlov et al.
Let find out how to change the values N and τ in formulas (3)–(4), if
probability variation of f (k) is low. Let a new distribution can be expressed
˜
f (k) = f (k) + ε(k)f (k), moreover, under the same norming the equality
as
ε(k)f (k) = 0 is fulfilled. Let introduce the variable E = sup |ε(k)|. Then we
k k
get the following estimation of distance between distributions in L1 norm:
ε = f − f˜ = f (k) − f˜(k) = |ε(k)| f (k) ≤ E. (5)
L1
k k
Let use (5) to get estimation of some differentiable function variation depend-
ing on average batch size in case E << 1:
(1)
δl = k f (k) − f˜(k) < k f (k) − f˜(k) = k |ε(k)| f (k) ≤ El(1) ;
k k k
(6)
∂u l(1)
δu l
(1)
= · δl(1) ≤ Eu ∂ ln u | (1) + o (E) .
∂ (l (1) ) ∂ ln l l=l
(7)
(2)
Eρ l − l (1)
b2
|δτ | ≤ b1 + .
2 · b
2 (1 − ρ) l (1)
1
δτ (ρ) 0, 106 + 0, 041 ln ρ (8)
0,27
τ (ρ) = ρ
(δf ) ;
ρ ∈ [0, 1; 0, 9].
v2 M −1 ρ b2 ρ2
N0 = ρ+ + 2 , (9)
2v1 b1 3 1 − ρ 2b1 1 − ρ
For basic distribution function Eq. (3) is considered as null approximation.
Step 3. Get the empirical batch size distribution function F (k) from measure-
ments. Let consider that empirical distribution, being investigated, is stationary,
F (k) is its distribution function.
Step 4. Calculate the distance between empirical and uniform distributions:
δf = sup |F (k) − k/M | . (10)
k
Step 5. Substituting the result of (10) in (8) we get estimation of average queue
length that corresponds to the following empirical distribution:
0,27
N/N0 ≈ 1 + 0, 2 (1 − ln ρ) (δf ) , ρ ∈ [0, 1; 0, 9]
(11)
Estimation for average waiting time variation is expressed in the same way.
Equation (11) is computationally much simpler than calculation of the gener-
ating function in accordance with (2), where L(z) is calculated through empirical
distribution of F (k). That is, firstly, rather difficult and, secondly, leads to cal-
culation errors that may exceed approximation inaccuracy for (11).
148 Y. Orlov et al.
For example, let consider that an empirical distribution of fn (k) is taken from a
general population f (k) that has geometric distribution with a parameter q on
the interval 1 ≤ k ≤ M . That means f (k) = 1−q M q
1−q k−1
.
We show you algorithm for average queue length calculation with geometric
distribution function.
Step 1. Basic batch size distribution:
f (k) = 1/8, 1 ≤ k ≤ M - uniform distribution.
Step 2. Average queue length N0 for the uniform distribution:
δf = 0, 35.
6 Conclusion
This paper presents the approach to estimation of the performance measures for
SIP server model with batch arrivals and vacations depending on the batch size
distribution. Investigation of this particular dependence was motivated by the
fact that the batch size distribution is not known as a general population and,
moreover, cannot be recognized as far as empirical evaluations of this popula-
tion are non-stationary. That is why approximate evaluation methods, that are
not associated with a specified functional class of mentioned distributions, are
of great significance and actuality. Therefore, the method that considers coeffi-
cients of the model parameters sensitivity to adjustment of the distance between
distribution functions seems to be efficient among nonparametric techniques.
This method may be used for non-stationery distributions when non-stationary
behavior is interpreted as definite variation of some basic distribution (for exam-
ple, uniform). This approach enables to circumvent technical difficulty coming
from absence of convergence theorem both for probability and the norm for
random variables being investigated.
Approach to Estimation of Performance Measures for SIP Server Model 149
Acknowledgment. This work was supported in part by the Russian Foundation for
Basic Research, projects No. 15-07-03051, 15-07-03608.
We thank Professor Konstantin Samouylov from Peoples’ Friendship University of
Russia for comments that greatly improved the paper.
References
1. Rosenberg, J., et al.: IETF RFC 3261 SIP: Session Initiation Protocol (2002).
http://www.ietf.org/rfc/rfc3261.txt
2. Rosenberg, J.: IETF RFC 5390 Requirements for Management of Overload in the
Session Initiation Protocol (2008). https://tools.ietf.org/html/rfc5390
3. Hilt, V., Noel, E., Shen, C., Abdelal, A.: IETF RFC 6357 Design Considerations
for Session Initiation Protocol (SIP) Overload, Control (2011). https://tools.ietf.
org/html/rfc6357
4. Gurbani, V., Hilt, V., Schulzrinne, H.: IETF RFC 7339 Session Initiation Protocol
(SIP) Overload Control (LBOC) (2014). https://tools.ietf.org/html/rfc7339
5. Noel, E., Williams, P.M.: IETF RFC 7415 Session Initiation Protocol (SIP) Rate
Control (RBOC) (2015). https://tools.ietf.org/html/rfc7415
6. Samouylov, K.E., Sopin, E.S.: On analysis of M[X]|G|1|r queuing system. Bul-
letin of Peoples’ Friendship University of Russia. Series Mathematics, Information
Sciences, Physics, no. 1, pp. 91–97 (2011)
7. Gaidamaka, Y., Pechinkin, A., Razumchik, R., Samouylov, K., Sopin, E.: Analysis
of M|G|1|R queue with batch arrivals and two hysteretic overload control policies.
Int. J. Appl. Math. Comput. Sci. 24(3), 519–534 (2014)
8. Gaidamaka, Y.V., Zaripova, E.R., Orlov, Y.N.: Analysis of the impact the batch
size distribution on parameters of the SIP-server queueing model with batch
arrivals. KIAM Preprint no. 27, pp. 1–16, Moscow (2015). http://library.keldysh.
ru/preprint.asp?id=2015-27
9. Gaidamaka, Y., Zaripova, E.: Comparison of polling disciplines when analyzing
waiting time for signaling message processing at SIP-server. In: Dudin, A., et al.
(eds.) ITMM 2015. CCIS, vol. 564, pp. 358–372. Springer International Publishing,
Switzerland (2015)
10. Gaidamaka, Y.V.: Model with threshold control for analyzing a server with an
SIP protocol in the overload mode. Autom. Control Comput. Sci. 47(4), 211–218
(2013)
11. Abaev, P., Gaidamaka, Y., Samouylov, K.E.: Modeling of hysteretic signaling load
control in next generation networks. In: Andreev, S., Balandin, S., Koucheryavy,
Y. (eds.) NEW2AN/ruSMART 2012. LNCS, vol. 7469, pp. 440–452. Springer, Hei-
delberg (2012)
12. Abaev, P., Gaidamaka, Y., Samouylov, K.E.: Queuing model for loss-based over-
load control in a SIP server using a hysteretic technique. In: Andreev, S., Balandin,
S., Koucheryavy, Y. (eds.) NEW2AN/ruSMART 2012. LNCS, vol. 7469, pp. 371–
378. Springer, Heidelberg (2012)
150 Y. Orlov et al.
13. Gaidamaka, Y., Pechinkin, A., Razumchik, R., Samouylov, K., Sopin, E.: Analysis
of M|G|1|R queue with batch arrivals and two hysteretic overload control policies.
Int. J. Appl. Math. Comput. Sci. 24(3), 519–534 (2014)
14. Abaev, P., Gaidamaka, Y., Pechinkin, A., Razumchik, R., Samouylov, K., Shorgin,
S.: Hysteretic control technique for overload problem solution in network of SIP
servers. Comput. Inform. 33(1), 218–236 (2014)
15. Shorgin, S., Samouylov, K., Gaidamaka, Y., Etezov, S.: Polling system with thresh-
old control for modeling of SIP server under overload. In: Advances in Intelligent
Systems and Computing, vol. 240, pp. 97–107 (2014)
16. Abhayawardhana, V.S., Babbage, R.: A traffic model for the IP multimedia subsys-
tem (IMS). In: Proceedings of 65th Vehicular Technology Conference, pp. 783–787
(2007)
17. Vishnevsky, V., Kozyrev, D., Semenova, O.: Redundant queuing system with unre-
liable servers. In: 2014 6th International Congress on Ultra Modern Telecommuni-
cations and Control Systems and Workshops (ICUMT), pp. 283–286 (2014)
18. Abaev, P.O., Gaidamaka, Y.V., Pechinkin, A.V., Razumchik, R.V., Shorgin, S.Y.:
Simulation of overload control in SIP server networks. In: Proceedings of the 26th
European Conference on Modelling and Simulation, ECMS 2012, Koblenz, Ger-
many, pp. 533–539 (2012)
19. Abaev, P.O., Pechinkin, A.V., Razumchik, R.V.: On analytical model for opti-
mal SIP server hop-by-hop overload control. In: Proceedings of the 4th Interna-
tional Congress on Ultra Modern Telecommunications and Control Systems (IEEE
ICUMT-2012), Saint-Petersburg, Russia, pp. 299–304 (2012)
20. Abaev, P., Gaidamaka, Y., Samouylov, K., Shorgin, S.: Design and software archi-
tecture of SIP server for overload control simulation. In: Proceedings of the 27th
European Conference on Modelling and Simulation (ECMS 2013), 27–30 May,
Aalesund, Norway, pp. 580–586 (2013)
21. Samouylov, K., Abaev, P., Gaidamaka, Y., Pechinkin, A., Razumchik, R.: Ana-
lytical modelling and simulation for performance evaluation of SIP server with
histeretic overload control. In: Proceedings of the 28th European Conference on
Modelling and Simulation (ECMS 2014), 27–30 May, Brescia, Italy, pp. 603–609
(2014)
22. Gaidamaka, Y., Pechinkin, A., Razumchik, R.: Time-related stationary character-
istics in queueing system with constant service time under hysteretic policy. In:
Proceedings of the 6th International Congress on Ultra Modern Telecommunica-
tions and Control Systems (ICUMT 2014), Russia, pp. 634–640 (2014)
23. Samouylov, K., Gaidamaka, Y., Talanova, M., Pavlotsky, O.: Simulation of SIP-
server with hysteretic input and loss-based overload control scheme. In: Proceed-
ings of the 6th International Congress on Ultra Modern Telecommunications and
Control Systems (ICUMT 2014), St. Petersburg, Russia, pp. 589–594 (2014)
24. Samouylov, K., Gaidamaka, Y., Abaev, P., Talanova, M., Pavlotsky, O.: Analytical
modeling of rate-based overload control with token bucket traffic shaping on client
side. In: Proceedings of the 29th European Conference on Modelling and Simulation
(ECMS 2015). Digitaldruck Pirrot GmbH, Albena, pp. 669–674 (2015)
The Synthesis of Service Discipline in Systems
with Limits
Taufik Aliev(B)
1 Introduction
One of the main characteristics of the functioning of computer systems of dif-
ferent classes and computer networks is the delay time of data processing and
transmission, which is named as residence time of the system in queuing sys-
tems used as models. The requirements for the quality of the functioning of such
systems are formulated in a variety of limits on the residence times in the differ-
ent classes of applications. For example, in data-processing systems, restrictions
apply to the mean residence time of the queries in the system (the average lim-
its), the excess of these limits does not lead to critical consequences. At the
same time, in information and control systems, existing in a circuit of automatic
control of technological equipment or mobile units, restrictions are imposed on
the probability of exceeding the set value of residence time (probabilistic limit)
which may lead to a sharp deterioration in the functioning of the system or
even to the exit its failure. In general, the quality requirements of the system
can be formulated as a combination of limits, where some classes have an aver-
age constraints, while others - probabilistic constraints. Applying limits ensuring
correct application of the priority strategies of process control and data process-
ing. Queuing system with one serving device and unlimited storage capacity,
which receives the inhomogeneous flow of requests to be processed with a given
capacity are widely used as models of such systems [1]. In this case, the problem
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 151–156, 2016.
DOI: 10.1007/978-3-319-30843-2 16
152 T. Aliev
2 Statement of a Problems
Load parameters for solving the problem of synthesis to a system with a non-
uniform flow of applications used: the number of classes of applications - H,
coming in from the intensities λ1 , . . . , λH and forms a simple flows; resource-
processing applications each class, asked for at least three factors: average values
θ1 , . . . , θH (commands or instructions executed by the processing of the appli-
cation of the corresponding class); coefficients of variation ν1 , . . . , νH and third
(3) (3)
initial moments θ1 , . . . , θH .
Suppose that the classes of applications are grouped so that the restrictions
on the residence times τu1 , . . . , τuH in the system for the first H1 classes are
defined in a probability, and for other grades - as mean limited to:
uh ≤ u∗h , (h = H1 + 1, H) (2)
where u∗h – allowable value of residence time in the application class h; δh∗ –
allowable probability of exceeding a specified limit u∗h ; Pr (τuh > u∗h ) – proba-
bility that the residence time in the application class h (h-request) τuh exceeds
the allowable value u∗h ; uh = M [τuh ] – average value (expectation) of residence
time of h-request in system.
Resident times τu1 , . . . , τuH depend on the system performance V and service
discipline. Obviously, for any restrictions to (1) and (2) can be performed at the
expense of the system performance. At the same time the best solution will be
the discipline which can perform with minimal restrictions performance.
The synthesis problem is solved service discipline in the classroom disciplines
with mixed priorities, which are described by a matrix of priorities Q = [qij (i, j =
1, . . . , H)], where qij set the priority of i -requests against j -requests: 0 – no
priority, 1 –relative priority (RP) 2 – absolute priority (AP) [1].
Thus, the problem of synthesis to systems with constraints is formulated as
follows: to find the discipline of service of requests with mixed priority given that
limits (1) and (2) are fulfilled with minimal performance V.
3 Calculated Identities
Resident time of h-requests (h = 1, H) in system τuh in common consists of the
waiting time of service start τxh and processing time τzh , including the waiting
time in the interrupted state: τuh = τxh + τzh . Then the expectation uh and the
(2)
second moment uh of resident time of h-requests:
(2) (2) (2)
uh = xh + zh ; uh = xh + 2xh zh + zh . (3)
The Synthesis of Service Discipline in Systems with Limits 153
(2) (2)
Values xh , zh and xh , zh are counted [1]:
H (2)
⎫
xh = i=1 r4 (i,k) λi bi
zh = bh ⎪
⎪
(2) (3) ; (1) ; ⎪
⎪
2(1−Rh ) (1−Rh )
H H
1−Rh
(2) H
⎪
⎪
(3) (2) ⎪
⎪
(2)
xh = i=1 r4 (i,k) λi bi
(2) (3) + i=1 r3 (i,k) λi bi
(2)
i=1 r4 (i,k) λi bi
(3)
⎬
3(1−Rh )2 (1−Rh ) 2(1−Rh )2 (1−Rh )2
H (2) H (2) (4)
i=1 r 2 (i,k) λi b i i=1 r 4 (i,k) λi bi ⎪
⎪
+ (2)
2(1−Rh )3 (1−Rh )
(3) ; ⎪
⎪
H ⎪
⎪
(2)
(2)
bh r (i,k) λ
(2)
bi ⎪
⎪
zh = (1) + i=1 1
(1) 3
i
, ⎭
(1−Rh )2 (1−Rh )
(n) (n)
where bi = θi /V n – n starting moment of service time of i -request
(g) H
(i = 1, H; n = 1, 2, 3); V – system performance; Rh = i=1 rg (i, h) λi bi –
partial total loads; rg (i, h) – coefficients with values 0 or 1, depending on
the values of the elements qih and qhi of priority matrix and allow appli-
cations to allocate classes i and h, with same priority class: r1 (i, h) =
0, 5 qih (qih − 1) – takes the value 1, if i -requests have absolute priority against
h-requests; r2 (i, h) = 0, 5 qih (3−qih ) – takes the value 1, if i -requests have relative
priority or absolute priority against h-requests; r3 (i, h) = 1−0, 5 qhi (3−2qih +qhi )
– takes the value 1, if i -requests have no priority or relative priority or absolute
priority against h-requests; r4 (i, h) = 1 + 0, 5 qhi (1 − qhi + qih ) – takes the value
0 only if h-requests have absolute priority against i -requests.
123
1011
Q= .
2002
3000
Class 2 request enters the system, when device is operating with the class
3 request, and the queue contains one or several class 1 requests, which cannot
interrupt the service of class 3 requests, as they have relative priority to class 3
(q13 = 1). This raises the following uncertainty. On the one hand, receiving the
class 2 request shall abort the service of class 3 request, because it has absolute
priority (q23 = 2) over class 3 requests. On the other hand, received class 2
request may not get service before class 1 requests are in a queue and have a
relative priority to class 2 requests (q12 = 1). This uncertainty leads to ambiguity
154 T. Aliev
uh ≤ ũh , (h = 1, H1 ). (5)
5 Conclusion
The proposed approach to the problem of the synthesis of service discipline with
mixed priorities for systems with complex limits on the residence times in the
different classes of applications allows for a specified quality of the system with
a minimum performance.
References
1. Aliev, T.I., Maharevs, E.: Queueing disciplines based on priority matrix. Sci. Tech.
J. Inf. Technol. Mech. Opt. 6(94), 91–97 (2014)
2. Aliev, T.I.: Approximation of probability distributions in queuing systems. Sci. Tech.
J. Inf. Technol. Mech. Opt. 2(84), 88–93 (2013)
Relative Navigation for Node of Wireless
Decentralized Network
1 Introduction
The development of wireless transmission is a rapid pace [1], along with the
navigation problem put their subscribers. In case of failure or absence of GPS
signals, or in areas without GSM coverage, or to improve positioning accuracy
relative navigation becomes necessary. This is especially true for decentralized
networks [2,3], when the nodes are mobile and move in space, such as sensor
networks or operative networks of tablet computers in places of emergencies [4].
There are two main methods for relative navigation subscriber in the network
GSM [5] method of obtaining the time (Time of Arrival – TOA); time difference
method (Enhanced Observed Time Difference, EOTD). The method of obtain-
ing the time (TOA) [6] similar to the GPS satellite navigation technology and is
based on the measurement of the delay in the shift of the frame with the signal
from the base station to the phone where the distance to the base station. To
determine the coordinates should be at least three simultaneous bearing. Calcu-
lations made by the operator of the triangulation algorithms. The time difference
method (EOTD) [7] measured delay time difference signals from the two closest
base stations, which is then converted to distance. Feature EOTD method is
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 157–166, 2016.
DOI: 10.1007/978-3-319-30843-2 17
158 D.A. Aminev et al.
the need to integrate the mobile terminal computing module, which also passed
the exact coordinates of the base stations, so blocks LMU base stations requires
several times less than the TOA. EOTD has spread in networks with CDMA
technology and is supported by some models of terminals to networks GSM.
According to some estimates, the accuracy of the method EOTD exceed TOA.
The disadvantage of these methods is the high accuracy of determining the
coordinates in a big city it is usually up to 400 m, in the regional center to a
kilometer in rural areas and along the routes of 15–20 km. In addition, both
methods require the installation of base stations LMU special module for cal-
culating position places. The methods are applicable only in the area of GSM
coverage.
Also known relative navigation method based on measurement of received
signal power [8,9]. In the method of approximation conducted a distant relation
of the type r1n characteristics, where n > 2. This takes into account attenuation
when passing the road and heterogeneity of the antenna pattern. This method
is not tied to coverage GSM, but, despite the simplicity of calculations obtained
by averaging feature provides low accuracy location.
A method of relative navigation unit wireless decentralized network based
on measurements of the received signal power from a distant approximation
characteristics dependent species r12 and taking into account the position error.
Fig. 1. Scheme geographical coordinate data from multiple fixed units (a), one movable
(b) and distant characteristic (c)
Since F2 (r) ≈ F1 (r) for r ≥ rc the conventional boundary near and far zones
antenna (Fig. 1c), and near-field problems locating antennas are not used, it is
possible to restrict the formula (1) for r ≥ rc . Furthermore, if the transmitter
and receiver are on the surface (the ground), then the reflection from it can be
neglected, and also to use the formula (1).
Action factors (b) and (c) generally is unknown, but it can be evaluated on
the basis of additional information about the terrain.
Factor (d) due to the heterogeneity of the diagrams of the transmitting and
receiving antennas. If you can change the orientation of the antenna in space, to
eliminate the influence of this factor requires information about their orientation.
If such information is not available, you must either request a measurement at a
specific antenna orientation (vertical, horizontal, etc.), or to carry out effective
filtering incoming data from the beacon. It is desirable that the beacons are
fairly evenly distributed over the neighborhood of the locator and the number
is sufficiently large.
Action factor (e) manifested usually at sufficient remoteness of the trans-
mitter and receiver from each other (several kilometers or more) and can be
accounted for using the weather data.
If the beacon is in the line of sight of the receiver, the distance to it is computed
directly from the distant characteristic patterns, taking into account:
−1 K1 PT
r = F (α) = = K1 R , (6)
α P
K2 −β
r = 10 20 . (7)
In this case, the locus is a circle of radius r, the center of which coincides
with the beacon. With two beacons locus is a pair of points (the intersection
of two circles), while the three — only one point (the intersection of the three
circles).
If you know the coordinates {(xk , yk )}m k=1 of the locator beacons are the
coordinates (x, y) of the system (the problem is solved multilateration):
⎧
⎪
⎨(x − x1 ) + (y − y1 ) = r1
2 2 2
.................. (8)
⎪
⎩
(x − xm )2 + (y − ym )2 = rm 2
.................. (9)
⎪
⎩
(φ − φm )2 cos2 θm + (θ − θm )2 = (rm /R)2
Fig. 2. The structure of the locus of the three beacons (a) and blur the “ideal” single
locus beacon (b)
1 K1
drk = (F −1 ) (αk )dαk = −
(1)
dαk (10)
2 αk3
Coordinates k-th beacon (xk , yk ) defined errors Δxk , Δyk , respectively (half
precision interval estimates), so the error location is the beacon
The total error in determining the distance to the beacon (half the thickness of
conventional “fuzzy ring”) equal to the sum of errors:
A rough estimate of the error location by this method can be obtained by aver-
aging the thickness of “fuzzy rings”:
1 1
m m
Δx := √ Δdk , Δy := √ Δdk . (20)
2m k=1 2m k=1
Relative Navigation for Node of Wireless Decentralized Network 163
3 Experiment
For the development of the technology has been developed relative navigation
layout (Fig. 3) and an experiment on the ground in the field (3b) for the embod-
iment of a decentralized network in accordance with 1b.
Based layouts are receiving and transmitting device based on TI CC1101
transceiver and microcontroller control ST STM32F0, a set of specially designed
programs for reception and transmission of navigation data. In addition, the
program shows the level RSSI, dBm. The means used are two laptops with
the operating system installed GNU/Linux, the two connected wireless trans-
ceiver module. The modules, developed by LLC “Open development” [10], are
small-sized wireless devices with a USB interface and a spiral antenna [11]. The
program runs on a laptop rf sender beacon and sends the coordinates of the
locator, rf reciever - runs on a laptop locator receives the data shows the level
of RSSI, calculates its own coordinates.
164 D.A. Aminev et al.
Bexp.
Bcalc.
a) b)
Fig. 3. The circuit layout (a) the location and the result on the ground (b)
When calculated in the first approximation consider the directivity pattern uni-
form, antenna gains believe unit, the accuracy of RSSI believe zero. Below is a
sequence of calculation.
(2) Find the distance r from the beacon by the formula (7) and distant charac-
teristic (Fig. 4).
(3) Solve the system of equations (6) for the unknowns φ and θ to obtain the
coordinates of the point estimate locator and obtain θ ≈ 55.83399999◦ , φ ≈
37.36739999◦ .
(4) To evaluate the error in determining the distance from the formulas (11) and
(16), and accept: Δβk = 0 and, therefore, Δrk = 0.
(5) Estimate the total error in determining the distance to the beacon by the
formula (19), where the error δk of GPS-location beacons given in Table 1.
(6) According to the formula (21) estimate error in determining the coordinates
of the locator an average thickness of
√ “fuzzy ring” (interval
√ estimation accu-
racy coordinate locator) Δx = Δd/ 2 ≈ 5 m, Δy = Δd/ 2 ≈ 5 m.
The resulting calculated position locator for GPS are as follows: latitude
θ ≈ 55.83399999◦ , longitude φ ≈ 37.36739999◦ , with an error of 5 m. The actual
discrepancy between calculated and experimental data was 0.0002◦ (0.72 ) in
latitude and longitude, which corresponds to 16 m. This discrepancy is due to the
chaotic traffic nodes network space and a random arrangement of their antennae,
which affects the radiation pattern and RSSI.
4 Conclusions
The proposed approach for the relative navigation allows to determine the geo-
graphical coordinates of the decentralized network nodes by the method of mul-
tilateration and evaluate the error location.
In the case of location on the plane in the presence of obstacles is recom-
mended to use the space-time filtering and temporal filtering eliminates the effect
of noise and space, ranging beacons in their accuracy and reliability, eliminates
the effect of remoteness and interference caused by the weight of the adaptive
system.
An indication of the importance of traffic nodes should be considered as a
change of the distance between them at the time of the signal and its processing.
This change in distance caused by the movement of nodes to each other and
the finite velocity of propagation and signal processing, causing a regular error
166 D.A. Aminev et al.
References
1. Tanenbaum, Je., Ujezeroll, D.: Komp’juternye seti.: 5-e izd. - SPb.: Piter, p. 960
(2012)
2. Azizov, R.F., Aminev, D.A., Ivanov, I.A., Uvajsov, S.U.: Organizacija svjazi na
osnove prioritetov dlja uluchshenija JeMS pri informacionnom obmene v decen-
tralizovannoj seti. Tehnologii jelektromagnitnoj sovmestimosti. 4(47), 5–8 (2013)
3. Azizov, R.F., Aminev, D.A., Uvajsov, S.U.: Programmnaja integracija jelementov
uzla decentralizovannoj seti. In: Innovative Information Technologies: Materials
of the International Scientific - Practical Conference. Part 3. - Praga, pp. 23–26
(2013)
4. Information Technology - Telecommunications and Information Exchange Between
Systems - Local and Metropolitan Area Networks - Specific Requirements - Part
15.4: Wireless Medium Access Control (MAC) and Physical Layer (PHY) Spec-
ifications for Low-Rate Wireless Personal Area Networks (WPANs) - Amend-
ment: Alternative Physical Layer Extension to Support the Japanese 950 MHz
Band/Note: supplement to ANSI/IEEE 802.15.4-2006 *Approved 28 July 2009
5. De Groote, A.: GSM Positioning Control, p. 13. University of Fribourg, Switzerland
(2005)
6. Yamasaki, R., Ogino, A., Tamaki, T., Uta, T., Matsuzawa, N., Kato, T.: TDOA
location system for IEEE 802.11b WLAN. In: Proceedings of Wireless Communi-
cations and Networking Conference, vol. 4, pp. 2338–2343 (2005)
7. Jonsson, D.K., Olavesen, J.: Estimated accuracy of location in mobile networks
using E-OTD. Master thesis in Information and Communication Technology, Grim-
stad, pp. 68 (2002)
8. Huang, X., Barralet, M., Sharma, D.: Accuracy of location identification with
antenna polarization on RSSI. In: Proceedings of the International MultiConfer-
ence of Engineers and Computer Scientists, IMECS 2009, vol. I. Hong Kong, 18–20
March 2009
9. Huang, T., Chen, Z., Xia, F., Jin, C., Liang, L.: A practical localization algorithm
based on wireless sensor networks. In: 2010 IEEE/ACM Int’l Conference on Green
Computing and Communications (GreenCom) & Cyber, Physical and Social Com-
puting (CPSCom), pp. 50–54 (2010)
10. http://open-dev.ru — site of the Corp. “Open development”
11. Aminev, D., Azizov, R., Uvaysov, S.: Recommendations for the choice of
antenna transceivers of decentralized self-organizing networks. V kn.: Innovacii
na osnove informacionnyh i kommunikacionnyh tehnologij: materialy mezhdunar-
odnoj nauchno-tehnicheskoj konferencii/Otv. red.: I. A. Ivanov; pod obshh. red.:
S. U. Uvajsova. M.: MIJeM NIU VShJe, pp. 480–481 (2013)
Algorithm of State Stationary Probability
Computing for Continuous-Time Finite Markov
Chain Modulated by Semi-Markov Process
Abstract. This paper presents the description of the method for com-
puting of state stationary probabilities for Markovian systems operating
in the random environment. These systems are described as semi-Markov
processes. The method generalizes some other results in this area.
1 Introduction
Lately a great attention has been granted to probabilistic models, taking a ran-
dom environment into consideration [4–8,10]. Generally it can be described by
the so-called Markov-modulated processes [1,2,5,11]. Here the random environ-
ment is represented as a continuous-time finite Markov chain X(t). For each
state of this chain, there exists its own “inner” continuous-time Markov chain
Y (t). The state set of all inner Markov chains is the same, but the transition
intensities are different.
We consider the following modification: random environment X(t) will be
presented by a semi-Markov process [6] instead of the Markov chain.
We present an efficient algorithm of stationary probabilities computing for
the states of process (X(t), Y (t)). This algorithm allows solving various partial
tasks, where a random environment has place [13].
The paper is organized as follows. The considered semi-Markov process X(t)
is described in the second section. There is considered an embedded Markov
chain that is built on the moments of states changing. The expectation S (i) of the
sojourn time is determined for each state i, and transition probabilities between
states. The third section contains description of inner continuous-time Markov
This research was financially supported by the Ministry of Education and Science
of the Russian Federation in the framework of the applied research project
№14.613.21.0020 of 22.10.2014 (RFMEFI61314X0020).
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 167–176, 2016.
DOI: 10.1007/978-3-319-30843-2 18
168 A.M. Andronov and V.M. Vishnevsky
chain Y (t). The expectations of the sojourn time in the states of Y (t) during time
S (i) is induced. Further the embedded discrete-time Markov chain (X(n), Y (n))
is considered, which is built on the time moments when semi-Markov process
X(t) changes own state: n = 1, 2, ... There a stationary distribution of the sate
probabilities is gotten. All these results allow calculate a stationary distribu-
tion of the sate probabilities for continuous-time chain (X(n), Y (n)) — Sect. 4.
A partial case, when continuous-time Markov chain takes place instead of the
semi-Markov process, is considered in the Sect. 5. The numerical example illus-
trates the suggested approach.
We denote μν,i an intensity of the output from state ν without changing the
state i of Markov chain: it is a sum of elements of the ν-th row of matrix M (i) :
m
(i) T
μν,i = Mν,ν ∗ . Let μi = μ1,i , ..., μm,i .
ν ∗ =1
Let Qν,ν ∗ (t) = P {Y (t) = ν ∗ |Y (0) = ν, X(τ ) = i, 0 ≤ τ ≤ t}, ν, ν ∗ = 1, ..., m.
(i)
(i)
The expression of matrix Q(i) (t) = Qν,ν ∗ (t) m×m is very well known [3,9,11].
(i) (i) T
and η1 , ..., ηm ; η̃1 , ..., η̃m are the rows of matrix N −1 .
(i) (i) (i) (i)
Let χ(i) = χ1 , ..., χm
Then
m
(i) (i) (i)
Q (t) =
(i)
ηζ xp(tχζ )η̃ζ , t ≥ 0. (4)
ζ=1
Note that expectation (1) can be got from (5). Actually, the first addend of
(i) (i)
the last row η1 η̃1 is the matrix with the same rows. Each row is a vector of
stationary distribution of the Markov chain. The sum of row elements equals
(i)
one. Other addends contain row-vectors {η̃ζ }. The sum of every row elements
equals zero. Therefore multiplying matrix E(U (i) ) from right by unit-vector we
get the vector-column which all elements equal the right of (1).
4 Embedded Markov-Chain
Now we consider an embedded discrete-time Markov chain (X(n), Y (n)), n =
0, 1, ..., which is built on the time moments when semi-Markov process X(t)
changes own state. Here X(n) is a new state of X(t) immediately after the n-th
transition. Note that state of Y (t) does not change at time moment t = S (i) .
(i,j) (i,j)
Let Rν,ν ∗ be the probability that the new state of X equals j: Rν,ν ∗ = P {Y (t) =
(i,j)
ν ∗ , X(S (i) ) = j|Y (0) = ν, X(τ ) = i, τ ∈ (0, S (i) )}. Let R(i,j) = Rν,ν ∗ m×m be
170 A.M. Andronov and V.M. Vishnevsky
∞ t
k
(i) (i) (i)
= ηζ η̃ζ λi,j (t) exp − Λi (τ )dτ + tχζ dt. (6)
ζ=1 0 0
An unique eigenvalue equals zero, let it is χ1 . Another are negative ones. The
eigenvector ν1 corresponding to χ1 , has identical elements, say c: ν• = (c, ..., c)T .
It allows to rewrite 6 as follows:
∞ t
(i) (i)
R (i,j)
= η1 η̃1 λi,j (t) exp − Λi (τ )dτ dt
0 0
∞ t
k
(i) (i) (i)
+ ηζ η̃ζ λi,j (t) exp − Λi (τ )dτ + tχζ dt. (7)
ζ=2 0 0
Let us remind, that h(i) = h̃(i) zi = h̃(i) ri − h̃(i) ed(m − 1) .
λi,j
ui,j = , i, j = 1, . . . , k. (12)
Λi
172 A.M. Andronov and V.M. Vishnevsky
r = rU. (13)
Further we suppose that distribution r is known, for example by this way. Let
us consider eigenvalues and eigenvectors of matrix U − I(k); let Ψ = (ψ1 , . . . , ψk )
be a matrix from the eigenvectors, Ψ −1 be corresponding inverse matrix with
rows ψ̃1 , . . . , ψ̃k . If the zero-eigenvalue has number i∗ , then all rows of matrix
ψi∗ ψ̃i∗ are vector r.
We have the following modification of formula (5):
k ∞
k
(i) (i) (i) (i) (i) 1
E(U (i) ) = ηζ η̃ζ exp(−Λi t + tχζ )dt = ηζ η̃ζ (i)
. (14)
ζ=1 0 ζ=1 Λ i − χζ
k ∞
k
(i) (i) (i) (i) (i) λi,j
R(i,j) = ηζ η̃ζ λi,j exp(−Λi t + tχζ )dt = ηζ η̃ζ (i)
. (15)
ζ=1 0 ζ=1 Λ i − χζ
(i)
Formulas (8) – (10) remain the same ones. Here hν = P {X = i, Y = ν}, i =
1, . . . , k, ν = 1, . . . , m, are stationary probabilities of the states immediately after
(i) (i)
a jump of Markov chain X(t). The vectors h(i) = h1 , . . . , hm , i = 1, . . . , k,
are calculated iteratively as it was described earlier.
Formula (11) for the stationary distribution of probabilities for the couple
(X, Y ) becomes simpler:
k −1
1
m
∗ (i)
P {X = i, Y = ν } = ri hν(i) E(Uν,ν ∗ ). (16)
i=1
λi ν=1
7 Numerical Example
Considered semi-Markov process X(t) has k = 3 states with the following tran-
sition intensities for the states.
A sojourn time of the 1st state has Weibull distribution with parameters a
and c. The corresponding rate is expressed as ([12], p. 400)
c x c−1
λV (x; a, c) = , x ≥ 0.
a a
The intensities of transition to states 2 and 3 are the following: λ1,2 (x) =
λV (x; 2, 2), λ1,3 (x) = λV (x; 3, 3), x ≥ 0.
Algorithm of State Stationary Probability Computing 173
A sojourn time of the 2nd state has Erlang distribution with parameter λ
and m. The corresponding rate is expressed as
1 m−1
λE(x; λ, m) = λ (λx) exp(−λx), x ≥ 0.
(m − 1)!
The intensities of transition to states 1 and 3 are the following: λ2,1 (x) =
λE(x; 3, 2), λ2,3 (x) = λE(x; 2, 2), x ≥ 0.
A sojourn time of the 3th state has Simpson distribution with parameter b.
The corresponding rate is expressed as
The intensities of transition to states 1 and 2 are the following: λ3,1 (x) =
λS(x; 4), λ3,2 (x) = λS(x; 6), 0 ≤ x < b.
The expectation of sojourn time of semi-Markov process in the state i is
calculated by formula (1). Vector of these values for different states is presented
below: T T
E(S) = E(S (1) ) · · · E(S (k) ) = 1.565 0.496 1.778 . (17)
Matrix P of one-step transition probabilities is calculated according to for-
mula (2): ⎛ ⎞
0 0.753 0.247
P = ⎝0.648 0 0.352⎠ .
0.743 0.257 0
The eigenvalues {νi } and eigenvectors of matrix P − I(3) are the following:
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
0 0.686 0.563 −0.213 0.592 0.592 0.592
ν = ⎝−1.609⎠ , Ψ = ⎝0.616 −0.793 0.789 ⎠ , Ψ −1
= ⎝1.419 −0.673 −1.447⎠ .
−1.391 0.386 0.230 −0.576 0.964 0.128 −1.917
We see that the zero eigenvalue has the first number (i∗ = 1), therefore
T T
Ψi∗ = Ψ1 = 0.686 0.616 0.386 , Ψ̃i∗ = Ψ̃1 = 0.592 0.592 0.592 .
The column of matrix Ψi∗ Ψ̃i∗ contain the same vector r = 0.406 0.365 0.229
that gives stationary probabilities of states:
r = 0.406 0.365 0.229 . (18)
Further we set m = 3 for the state number of the inner Markov chain Y (t).
The m × m-matrices of transition intensities for Y (t) and for different state i of
the random environment are the following:
174 A.M. Andronov and V.M. Vishnevsky
⎛ ⎞
0 0 0.3
= ⎝0.1 0 0.1⎠ ,
(1)
M (1) = Mν,ν ∗
0.3 0.1 0
⎛ ⎞
0 0.2 0.3
= ⎝0.2 0 0.1⎠ ,
(2)
M (2) = Mν,ν ∗
0.2 0.1 0
⎛ ⎞
0 0.4 0.1
= ⎝0.3 0 0.2⎠ .
(3)
M (3) = Mν,ν ∗
0 0.2 0
(i)
Formula (4) gives the nonstationary probabilities of Y (t)’s states Qν,ν ∗ (t) =
P {Y (t) = ν ∗ | Y (0) = ν, X(τ ) = i, 0 ≤ τ ≤ t}, ν, ν ∗ = 1, . . . , m. For example, we
have the following eigenvalues and eigenvectors for the second state:
⎛ ⎞ ⎛ ⎞
0 −0.577 −0.722 0.160
ν = ⎝−0.741⎠ , N = ⎝−0.577 −0.143 −0.811⎠ .
−0.459 −0.577 0.677 0.563
Note that the element sum of each row equals 0.497, it is the expectation of
sojourn time at state i = 2 of X(t) during single visit (see formula (1)).
Now we consider two-dimensional Markov chain (X(t), Y (t)) , t > 0. For
example, we present the matrix (6) for the states i = 2, j = 1 and j = 3 of
embedded Markov chain:
⎛ ⎞ ⎛ ⎞
0.522 0.051 0.043 0.281 0.029 0.042
= Rν,ν ∗ = ⎝0.050 0.567 0.017⎠ , R = Rν,ν ∗ = ⎝0.028 0.306 0.017⎠ .
(2,1) (2,1) (2,3) (2,3)
R
0.050 0.029 0.568 0.028 0.017 0.307
Let us note that element sum of each row for the first matrix equals 0.648
and the same for the second matrix equals 0.352.
Further we illustrate the above described iterative procedure (8) using matri-
ces. Matrices W (2,1) , W (2,3) and vectors w(2,1) , w(2,3) are the following:
0.522 0.051 0.281 0.029
W (2,1) = , W (2,3) = ,
0.050 0.567 0.028 0.306
w(2,1) = 0.050 0.029 , w(2,3) = 0.028 0.017 .
Algorithm of State Stationary Probability Computing 175
(i)
The stationary probabilities hν = P {X = i, Y = ν} for the embedded
Markov chain are calculated as h(i) = h̃(i) ri − h̃(i) ed(m − 1) . They are pre-
sented by matrix
⎛ (1) ⎞ ⎛ ⎞
h 0.111 0.130 0.165
H̃ = ⎝h(2) ⎠ = ⎝0.105 0.097 0.163⎠ (20)
h (3) 0.092 0.134 0.003
Addition of the rows and the columns give us the marginal probability dis-
tribution for the states of semi-Markov process X(t) and Markov chain Y (t):
0.520
P {X = i} = 0.333 0.360 0.308 , P {Y = ν} = 0.148 .
0.333
8 Conclusion
The described method gives a uniform approach to computing of the state sta-
tionary probabilities for Markovian systems operating in random environment.
The last are described as semi-Markov processes. The method generalizes some
another results in this area.
References
1. Andronov, A.M.: Markov-modulated birth-death processes. Autom. Control Com-
put. Sci. 45(3), 123–132 (2011)
2. Andronov, A., Gertsbakh, I.B.: Signatures in Markov-modulated processes. Stoch.
Models 30, 1–15 (2014)
3. Bellman, R.: Introduction to Matrix Analysis. Mcgraw-Hill Book Company Inc.,
New York (1969)
4. Kim, C.S., Dudin, A., Klimenok, V., Khramova, V.: Erlang loss queueing system
with butch arrivals operating in a random environment. Comput. Oper. Res. 36(3),
674–697 (2009)
5. Kim, C.S., Klimenok, V., Mushko, V., Dudin, A.: The BMAP/PH/N retrial queue-
ing system operating in Markovian random environment. Comput. Oper. Res.
37(7), 1228–1237 (2010)
6. Grabski, F.: Semi-Markov Processes: Application in System Reliability and Main-
tenance. Elsevier, Amsterdam (2014)
7. Kopoci’nska, I., Kopoci’nski, B.: On system reliability under random load of ele-
ments. Aplicationes Math. XVI(1), 5–14 (1980)
8. Kopoci’nska, I.: The reliability of an element with alternating failure rate.
Zastosowania Matematki (Aplicationes Math.) XVIII(2), 187–194 (1984)
9. Kijima, M.: Markov Processes for Stochastic Modelling. Chapman & Hall, London
(1997)
10. Leri, M.: Forest fire model for configuration graphs in a random environment. In:
Rasch, D., Melas, V., Pilz, J., Moder, K., Spangl, B. (eds.) Eighth International
Workshop on Simulation. Book of Abstracts. Institute of Applied Statistics and
Computing, University of Natural Resources and Life Sciences, Vienna, pp. 125–
126 (2015). ISBN 978-3-900932-28-2
11. Pacheco, A., Tang, L.C., Prabhu, N.U.: Markov-Modulated Processes & Semire-
generative Phenomena. World Scientific, New Jersey (2009)
12. Sleeper, A.: Six Sigma Distribution Modeling. McGraw Hill, New York (2007)
13. Vishnevsky, V.M., Semenova, O.V., Sharov, S.Y.: Modeling and analysis of a hybrid
communication channel. Autom. Remote Control 72, 345–352 (2013)
Approximate Description of Dynamics
of a Closed Queueing Network
Including Multi-servers
Svetlana Anulova(B)
1 Introduction
1.1 Review of Investigated Contact Centers Models
In the last ten years an extensive research in mathematical models for telephone
call centers has been carried out, cf. [2–6,8–17]. The object has been expanded to
more general customer contact centers (with contact also made by other means,
such as fax and e-mail). One of important relating questions is the dynamics
of multi-server queues with a large number of servers. In order to describe the
object efficiently the state of the model must include: (1) for every customer in
the queue the time that he has spent in it and (2) for every customer in the
multi-server the time that he has spent after entering the service area, that is
being received by one of the available servers.
This work was supported by RFBR grant No. 14-01-00319 “Asymptotic analysis of
queueing systems and nets”.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 177–187, 2016.
DOI: 10.1007/978-3-319-30843-2 19
178 S. Anulova
We suggest here a more suitable model for contact centers. The number of cus-
tomers is fixed. Customers may be situated in two states: normal and failure.
There is a multi-server which repairs customers in the failure state. The repair
time/the time duration of a normal state is a random variable, independent
and identically distributed for all customers. Now “the arrival process” in the
multi-server does not correspond to that of the previous G/GI/s + GI model.
For a large number of customers and a suitable number of servers calculate the
number of current failures, so much as an approximation. This is a continuation
of our work [1], where a single multi-server in a network was functioning.
We confine ourselves to a discrete time model. W. Whitt has written a very
interesting seminal article [15], in a simple discrete case. About 150 authors have
cited it and made generalizations to the continuous time. But their results do
not enclose Whitt’s discrete time ones. Walsh Zuñiga [12] with his results most
close to discrete time admits only discrete time service but his arrival process is
continuous.
In Whitt’s article [15] the idea of the convergence proof is true and very lucid,
and the proof is clearly presented. But Whitt has not covered all cases in his
proofs, and at that Whitt does not accurately point to it. Only the main case
is examined: the number of serviced customers is not zero and does not exceed
the number of customers in the queue. Also when he proves convergence of the
number of customers being served for a given time b(t, k) in (6.33) he omits the
case bs (n − 1, k − 1) tends to zero. We have transferred his proof technique to
our new network model and filled all these misses.
the normal state), the time they service a customer has distribution G1 . Multi-
server 2 (further denoted MS2) consists of sn n servers with a number sn ∈ (0, 1)
(for the customers in the failure state), the time they service a customer has
distribution G2 . The distributions G1 , G2 are discrete: they are concentrated on
{1, 2, . . .}. Service times are independent for both servers and all customers. We
will investigate the behavior of the net as n → ∞, namely, we shall establish a
stochastic-process fluid limit. It will be done only in a special case: discrete time
t = 0, 1, 2, . . .
We begin with a simple example of functioning of this network.
with bin (t, k) being the number of customers in the multi-server i at the moment t
who have spent there time k, i = 1, 2, and qn (t, k) being the number of customers
in the queue at the moment t who have been there precisely for time k. bin (t, k)
may also be interpreted as the number of busy servers at time t in the multi-
server i that are serving customers that have been in service precisely for time
k, i = 1, 2.
At the same time moment t ∈ {1, 2, . . .} multiple events can take place, so
we have to specify their order.
We must create a fictitious queue for the MS1—in fact this multi-server is so
large (n servers), that any customer of the whole quantity n trying to enter the
MS1 at once finds a free server in it.
At the time moment t the parameters b1 , b2 , q are taken from the previous
time t − 1 and processed to the current situation.
For both multi-servers:
– third, waiting customers in queue move into service of the multi-server accord-
ing to Assumption 1.
Customers enter service in MS2 whenever a server is available, so that the
system is work-conserving; i.e. we assume that Qn (t) = 0 whenever Bn2 (t) < sn n,
and that Bn2 (t) = sn n whenever Qn (t) > 0, t = 0, 1, 2, . . .. This condition can be
summarized by the equation
Notations. Let σni (t) be the number of service completions in MSi at time
moment t = 1, 2, . . ., i = 1, 2. Denote for k = 1, 2, . . . Gi;c (k) = 1 − Gi (k) and
g i (k) = Gi (k) − Gi (k − 1), i = 1, 2. Symbol ⇒ means convergence of the network
state characteristics to a constant in probability as the index n denoting the
number of cusomers tends to infinity.
Theorem 1 (The Discrete-Time Fluid Limit). Suppose that for each n, the
system is initialized with workload characterized by nonnegative-integer-valued
stochastic processes
satisfying
Bn1 (0) + Bn2 (0) + Qn (0) = n, (1)
Bn2 (0) ≤ sn n, and (sn n − Bn2 (0))Qn (0) =0 (2)
for each n w.p.1. Suppose that sn → s ∈ (0, 1) and
Then, as n → ∞,
bin (t, k)
⇒ bi (t, k), i = 1, 2, (5)
n
qn (t, k)
⇒ q(t, k), (6)
n
σni (t)
⇒ σ i (t), i = 1, 2, (7)
n
for each t ≥ 1 and k ≥ 0, where (b1 , b2 , q, σ 1 , σ 2 ) is a vector of deterministic
functions (all with finite values).
Approximate Description of Dynamics of a Closed Queueing Network 181
with
B 1 (t), B 2 (t), Q(t) ≥ 0, B 1 (t) + B 2 (t) + Q(t) = 1, (10)
B (t) ≤ s, and (s − B (t))Q(t) = 0.
2 2
(11)
The evolution of the vector (b , b , q, σ , σ )(t), t = 0, 1, 2 . . . , proceeds with steps
1 2 1 2
Case 2. B 2 (t − 1) < s. In this second case, after the time moment t − 1 asymp-
totically all servers are not busy so that there is no queue. As in the first case,
Eqs. (12), (13), and (14) hold. Instead of (15),
b2 (t, 0) = min{s − B 2 (t − 1) + σ 2 (t), σ 1 (t)}. (21)
Then,
q(t, k) = 0 f orall k > 0 and q(t, 0) = σ 1 (t) − b2 (t, 0). (22)
182 S. Anulova
Proof. For t = 0 conditions (3) and (4) imply the convergence presented in (8)
and (9), and conditions (1), (2) provide properties (10), (11). Thus the proof of
the Theorem is reduced to the following Lemma:
Lemma 1. Suppose for a given t ∈ {1, 2, . . .} Eqs. (5), (6) hold for t − 1. Then
all the statements of the Theorem hold for t.
Corollary 1. Condition (4) holds not only for time t = 0, but for any time
t = 1, 2, . . ., i.e., for each t, > 0 and η > 0, there exists an integer k0 such that
for n = 1, 2, . . .
∞
∞
bin (t, k) qn (t, k)
P( > ) < η, i = 1, 2, and P( > ) < η. (25)
n n
k=k0 k=k0
bin (t − 1, k − 1)
⇒ bi (t − 1, k − 1),
n
then
σni (t, k) g i (k)
⇒ bi (t − 1, k − 1) i;c , k = 1, 2, . . . , i = 1, 2,
n G (k − 1)
bin (t, k) Gi;c (k)
⇒ bi (t, k) = bi (t − 1, k − 1) i;c , k = 1, 2, . . . , i = 1, 2.
n G (k − 1)
Approximate Description of Dynamics of a Closed Queueing Network 183
Proof. Set i = 1, 2. For each n, t, k ≥ 1, we can represent σni (t, k) and bin (t, k) as
random sums of IID Bernoulli random variables; in particular,
bin (t−1,k−1)
σni (t, k − 1) = Xi (26)
i=1
and
bin (t−1,k−1)
bin (t, k) = (1 − Xi ), (27)
i=1
where Xi assumes the value 1 if the ith customer among those in service at time
moment t − 1 that have been in the system for time k − 1 is served at time
moment t, and assumes the value 0 otherwise. Thus, Xi , i ≥ 1, is a sequence of
IID random variables with
g i (k)
P(X1 = 0) = , k = 1, 2, . . . , i = 1, 2.
Gi;c (k − 1)
Apply now Appendix Lemma 5.
Corollary 2. If for given t ∈ {1, 2, . . .}, i ∈ {1, 2}
bin (t − 1, k − 1)
⇒ bi (t − 1, k − 1), k = 1, 2, . . . ,
n
then limit (7) holds:
σni (t)
⇒ σ i (t). (28)
n
Proof. This is guaranteed by Lemma 2.
Now we shall calculate the queue after it has filled the free servers. For a given
n ∈ {1, 2, . . .} we move σn2 (t) customers from the end of queue, that is, the
customers who have spent the longest time in the queue (if there are fewer
customers in the queue, we move they all).
Define an intermediate queue qn :
qn (t, 0) = σn1 (t), qn (t, k) = qn (t − 1, k − 1), k = 1, 2, . . . (29)
Obviously,
if σn2 (t) = 0 then qn (t, k) = qn (t, k), k = 0, 1, . . . , (30)
∞
if σn2 (t) ≥ qn (t, k) then qn (t, k) = 0, k = 0, 1, . . . , (31)
k=0
∞
if 0 < σn2 (t) < qn (t, k) then with (32)
k=0
∞
cn (t) = min{i ∈ {0, 1, . . .} : qn (t, k) ≤ σn2 (t)}, (33)
k=i
184 S. Anulova
⎧
⎪
⎪0 for k ≥ cn (t),
⎪
⎨ ∞
qn (t, k) = qn (t, i) − σn2 (t) for k = cn (t) − 1,
⎪
⎪i=c (t)−1
⎪ n
⎩q (t, k) for k < c (t) − 1.
n n
qn (t, k)
⇒ q (t, k), k = 1, 2, . . .
n
It is easy to understand other presentations of qn and q:
Lemma 4. For i = 0, 1, 2, . . .
∞
∞
qn (t, k) = ( qn (t, k) − σn (t)) ∨ 0, (34)
k=i k=i
∞
∞
q(t, k) = ( q (t, k) − σ(t)) ∨ 0. (35)
k=i k=i
It follows straightforward:
∞
∞
qn (t, k) qn (t, k) σn (t)
=( − ) ∨ 0. (36)
n n n
k=i k=i
qn (t, k) −
a−
n (k) = , a (k) = q(t, k) and θ = σ(t),
n
and using Remark 1 and Corollary 2, we obtain convergence in Eqs. (6), (7).
2 Appendix
1
ξn
lim Xi = ξP(X1 = 1). (37)
n→∞ n
i=1
Approximate Description of Dynamics of a Closed Queueing Network 185
ξn
Proof. As Xi ≤ ξn , n = 1, 2, . . ., the proposition is evident for ξ = 0. If
i=1
ξ > 0, then limn→∞ ξn = ∞ and (weak LLN)
ξn
1
ξn
ξn Xi
Xi = i=1
−−−−→ ξEX1 = ξP(X1 = 1).
n i=1 n ξn n→∞
Proof. Whitt has introduced this Assumption and derives from it partially the
statement of the Lemma for i = 0 without proof (see the first paragraph in the
proof of Theorem 6.1 [15]).
Lemma 7. Suppose that Assumption 2 holds and nonnegative random variables
θn , n = 1, 2, . . . , with deterministic limit θ, θn ⇒ θ, are given.
Define new sequences1 a− n , n ∈ {1, . . .}, a
−
by the following equations
sequences: for i = 0, 1, 2, . . .
∞
∞
a−
n (k) = ( an (k) − θn ) ∨ 0, (41)
k=i k=i
∞
∞
a− (k) = ( a(k) − θ) ∨ 0. (42)
k=i k=i
These a− −
n , n ∈ {1, . . .}, a sequences satisfy:
2
∞
∞
a− −
n (k) ⇒ a (k), k = 0, 1, 2, . . . , and a−
n (k) ⇒ a− (k) . (43)
k=0 k=0
1
They describe queues after customers leave them and go to the emptied servers.
2
Even starting from an arbitrary k.
186 S. Anulova
what is identical to
∞
∞
a−
n (k) ⇒ a− (k). (46)
k=i k=i
Finally, for i = 0, 1, 2, . . .
∞
∞
∞
∞
a−
n (i) = a−
n (k) − a−
n (i) ⇒ a− (k) − a− (k) = a− (k). (47)
k=i k=i+1 k=i k=i+1
3 Conclusion
We have investigated a model without abandonment in the queue, although the
necessary details of the customers age in the queue are provided. Since such a
behavior in the queue is universally recognized, we intend to consider it next.
As customers in a closed network cannot abandon it, probably we shall choose
instead a similar version of “nonpersistent customers”, see [7].
References
1. Anulova, S.V.: Age-distribution description and “fluid” approximation for a net-
work with an infinite server. In: International Conference “Probability Theory and
its Applications”, Moscow, pp. 219–220. 26–30 June 2012. (M.: LENAND)
2. Brown, L., Gans, N., Mandelbaum, A., Sakov, A., Shen, H., Zeltyn, S., Zhao, L.:
Statistical analysis of a telephone call center: a queueing-science perspective. J.
Am. Stat. Assoc. 100(469), 36–50 (2005)
3. Dai, J., He, S.: Many-server queues with customer abandonment: a survey of diffu-
sion and fluid approximations. J. Syst. Sci. Syst. Eng. 21(1), 1–36 (2012). http://
dx.doi.org/10.1007/s11518-012-5189-y, and http://link.springer.com/article/10.
1007/s11518-012-5189-y
4. Gamarnik, D., Goldberg, D.A.: On the rate of convergence to stationarity of the
M/M/n queue in the Halfin-Whitt regime. Ann. Appl. Probab. 23(5), 1879–1912
(2013)
5. Gamarnik, D., Stolyar, A.L.: Multiclass multiserver queueing system in the Halfin-
Whitt heavy traffic regime: asymptotics of the stationary distribution. Queueing
Syst. 71(1–2), 25–51 (2012)
Approximate Description of Dynamics of a Closed Queueing Network 187
6. Kang, W., Pang, G.: Equivalence of fluid models for Gt/GI/N+GI queues. ArXiv
e-prints, February 2015. http://arxiv.org/abs/1502.00346
7. Kang, W.: Fluid limits of many-server retrial queues with nonpersistent cus-
tomers. Queueing Systems (2014). http://gen.lib.rus.ec/scimag/index.php?s=10.
1007/s11134-014-9415-9
8. Kaspi, H., Ramanan, K.: Law of large numbers limits for many-server queues. Ann.
Appl. Probab. 21(1), 33–114 (2011)
9. Koçağa, Y.L., Ward, A.R.: Admission control for a multi-server queue with aban-
donment. Queueing Syst. 65(3), 275–323 (2010)
10. Pang, G., Talreja, R., Whitt, W.: Martingale proofs of many-server heavy-traffic
limits for Markovian queues. Probab. Surv. 4, 193–267 (2007). http://www.emis.
ams.org/journals/PS/viewarticle9f7e.html?id=91&layout=abstract
11. Reed, J.: The G/GI/N queue in the Halfin-Whitt regime. Ann. Appl. Probab.
19(6), 2211–2269 (2009)
12. Zuñiga, A.W.: Fluid limits of many-server queues with abandonments, general ser-
vice and continuous patience time distributions. Stochast. Process. Appl. 124(3),
1436–1468 (2014)
13. Ward, A.R.: Asymptotic analysis of queueing systems with reneging: A survey of
results for FIFO, single class models. Surv. Oper. Res. Manage. Sci. 17(1), 1–14
(2012). http://www.sciencedirect.com/science/article/pii/S1876735411000237
14. Whitt, W.: Engineering solution of a basic call-center model. Manage. Sci. 51(2),
221–235 (2005)
15. Whitt, W.: Fluid models for multiserver queues with abandonments. Oper. Res.
54(1), 37–54 (2006). http://pubsonline.informs.org//abs/10.1287/opre.1050.0227
16. Xiong, W., Altiok, T.: An approximation for multi-server queues with deterministic
reneging times. Ann. Oper. Res. 172, 143–151 (2009). http://link.springer.com/
article/10.1007/s10479-009-0534-3
17. Zhang, J.: Fluid models of many-server queues with abandonment. Queue-
ing Syst. 73(2), 147–193 (2013). http://link.springer.com/article/10.1007/
s11134-012-9307-9
Probability Characteristic Algorithm
of Upstream Traffic in Passive Optical Network
1 Introduction
Each ONU uses a reconfigurable laser and can transmit upstream traffic
in the selected range of wavelengths. The dynamically allocating wavelengths
mechanism allows increasing the system capacity, providing highly scalable. As
well as WDM-TDMA PON provides new users and required quality of service
level through existing frequency plan.
Parameters Description
L Number of optical network units
W , W ≤ L Number of upstream traffic wavelengths
κl , l = 1, L l-requests arrival intensity to assign a wavelength
νl , l = 1, L ONUl transition rate from the ON- to OFF-state
L
Ω := {n|n• ∈ {0, 1, . . . , W }, W ≤ L}, n• := 1T n = ni . (2)
i=1
Then,
W
L
|Ω| = . (3)
w
w=0
Probability Characteristic Algorithm of Upstream Traffic in PON 191
The state subspace when ONUl is in the ON-state or can pass into the ON-
state at the activation moment
Then,
W −1
L−1
|Ωl,ON | = 2 . (5)
w
w=0
Then,
L−1
|Ωl,OFF | = . (7)
W
The queue system functioning is described by a stationary Markov process
X(t) := (Xl (t))l=1,L , t > 0. Xl (t) = 1 means, that ONUl is in the ON-state,
Xl (t) = 0 means, that ONUl is in the OFF-state.
Theorem 1. The stationary Markov process X(t), t > 0, has got the stationary
probability distribution,which does not depend on the starting probability distrib-
ution and is multiplicative
L
1
L
κl
−1
p(n) = G anl l , n ∈ Ω, G = = anl l , al := , l = 1, L. (8)
p(0) νl
l=1 n∈Ω l=1
W
αl = 1 − G−1 g−l (W ) , G = g (L, w) , (10)
w=0
1, w = 0,
g−l (w) = (11)
g (L, w) − al g−l (w − 1) , w = 1, W ,
⎧
⎨ 0, l = 0, w = 1, W ,
g(l, w) = 1, l = 0, L, w = 0, (12)
⎩
g (l − 1, w) + al g (l − 1, w − 1) , l = 1, L, w = 1, W ,
where al := κνll , l = 1, L.
αl , l = 1, L, is one of the key performance factors of the Network I.
αl μk , l = 1, L, k = 1, K. (13)
The queue system scheme of the upstream traffic transmitting from ONU to
OLT is shown in the Fig. 3.
The queue system in the Fig. 3 is described using the following parameters.
Probability Characteristic Algorithm of Upstream Traffic in PON 193
K
S := {M|0 ≤ bk ml,k ≤ Rl , l = 1, L}. (15)
k=1
K
Sl,k := {M ∈ S| bk ml,k ≤ Rl − bk }, l = 1, L, k = 1, K. (16)
k=1
K
S̄l,k = S\Sl,k = {M ∈ S| bk ml,k > Rl − bk }, l = 1, L, k = 1, K. (17)
k=1
Theorem 2. The stationary Markov process Y(t), t > 0, has got the stationary
probability distribution,which does not depend on the starting probability distrib-
ution and is multiplicative
L
1
K
p(M) = G−1
m
ml,• ρl,kl,k ,
α
l=1 l k=1
1 L
1 ml,k
K
G= = ml,• ρl,k , (18)
p(0) α
M∈S l=1 l k=1
λl,k K
M ∈ S, ρl,k := , k = 1, K, ml,• := ml,k , l = 1, L.
μk
k=1
1
R
πk = g(K, r), k = 1, K. (22)
G
r=R−bk +1
Parameters Value
K 2
L 16
W 0, 16
Rl ,l = 1, L 18
T
b (1;2)
ρ11 , ρ12 0.5, 0.25
ρlk , l = 2, L, k = 1, K 1
1
0,9
0,8
0,7
0,6
0,5
0,4
0,3
0,2
0,1
0 W
0 1 2 3 4 5 6 7 8 9 10
Fig. 7. The chart of the blocking probabilities π11 and π12 from W .
8 Conclusion
In this paper, the mathematical model of the upstream multiservice traffic trans-
mitting from L ONU to OLT in the Network I is presented. The algorithm for
calculating of the blocking probability is proposed. In addition, there is the
numerical analysis example in which the role of W is shown to select the opti-
mal functioning behavior of the ONU.
The authors recommend the use of the concepts described in this paper
for the algorithm development to calculate of the blocking probability for the
upstream priority traffic transmitting model in the Network I.
References
1. Basharin, G.P.: Lectures on mathematical teletraffic theory. The 3rd publication.
PFUR, Moscow (2009)
2. Listvin V.N., Treschikov V.N.: DWDM-systems. The 2nd publication. TECH-
NOSPHERA, Moscow (2015)
3. Efimushkin, V.A., Savangukov, I.M.: Resource Allocation in Optical Transport
Networks. CRIC, Moscow (2010)
4. Ramaswami, R., Sivarajan, K.N., Sasaki, G.H.: Optical Networks A practical Per-
spective, 3rd edn. Morgan Kaufmann Publishers Inc., San Francisco (2010)
5. Mukherjee, B.: Optical WDM Networks. Springer, New York (2006)
6. Siva Ram Murthy, C., Gurusamy, M.: WDM Optical Networks: Concepts Design
and Algorithms. Prentice Hall PTR, Upper Saddle River (2002)
7. Basharin, G.P., Gaidamaka, Y.V., Rusina, N.V.: Probability characteristic compu-
tation algorithm of ONUs functioning in PON. Vestn. PFUR 2, 28–32 (2015)
8. Naymov, V., Samyilov, Y., Yarkina, N.V.: Teletraffic theory of multiservice net-
works. Monography. PFUR, Moscow (2008)
9. Basharin, G., Rusina, N.: Multirate loss model for optical network unit in passive
optical networks. In: Vishnevsky, V., Kozyrev, D., Larionov, A. (eds.) Distrib-
uted Computer and Communication Networks. Communications in Computer and
Information Networks, vol. 279, pp. 219–228. Springer, Cham (2014)
10. Basharin, G.P.: Maintenance of two streams with relative priority for a fully acces-
sible system with a limited waiting space. In: Proceedings of the AS of the USSR,
Technical Cybernetics, vol. 2, pp. 72–86 (1967)
11. Busen, J.P.: Computational algorithms for closed queueing networks with expo-
nential servers. Commun. ACM 16, 527–531 (1973)
12. Basharin, G.P., Gaidamaka, Y.V., Samouylov, K.E.: Mathematical theory of tele-
traffic and its application to the analysis of multiservice communication of next
generation networks. Autom. Control Comput. Sci. 47(2), 62–69 (2013)
Redundant Distribution of Requests Through
the Network by Transferring Them Over
Multiple Paths
1 Introduction
Reliability and fault-tolerance of distributed computing systems are convention-
ally achieved by applying duplicate computing and communication resources.
Owing to the use of redundant network components, requests are distributed
adaptively among network links and computing nodes that are in order at the
moment. Such approach relies on the comprehensive system design and assumes
partial loss of requests due to the network failures. In the critical systems, how-
ever, any loss of requests may cause the whole logic of computations to fail,
implying considerable consequences in terms of money, safety, and others.
The study focuses on distribution of requests through highly reliable networks
by using the mechanisms of multipath routing. Routing is referred to as the
choice of the route that meets specified requirements for the current dataflow.
In case of single-path routing, requests are sent over a single path, the best one
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 199–207, 2016.
DOI: 10.1007/978-3-319-30843-2 21
200 V.A. Bogatyrev and S.A. Parshutina
from a certain point of view. In case of multipath routing, there exists a set of
appropriate paths from the sender (the source of requests) to their receiver and
each request is transmitted over several paths simultaneously.
The use of multipath routing mechanisms serves to avoid overloads and uti-
lize network resources in the optimal way. The considerable part of the models
that involve those routing techniques addresses the problem of network load bal-
ancing, like [1]. Nevertheless, multipath routing is also discussed in relation to
network survivability, recovery, security, and energy efficiency [2].
Employing multipath routing techniques suggests that data packets are dis-
tributed over multiple paths either evenly or on grounds of one or more criteria.
In the former case, Equal-Cost Multi-Path (ECPM) routing strategies are imple-
mented, which means that there are no less than two routes having equal min-
imum cost of data transmission. The decision about the next hop, or the path
to forward a packet along, is taken on the basis of a number of algorithms. For
example, a Hash-Threshold algorithm involves performing a hash over the packet
header fields and utilizing the resulting threshold value, while the Round-Robin
principle implies choosing the least recently used next-hop [3]. In the latter case,
packets are distributed over paths according to a certain criterion or a combi-
nation of several criteria. It can be bandwidth, requests’ time in the system, or
some other metrics, which serves as a basis for calculating special values. Those
values, or indices, are assigned to all possible routes and thus rank them by pref-
erence, depending on the metrics. One more criterion that deserves attention
is fault statistics, which underlies the idea to prefer more reliable paths to less
reliable ones.
Contemporary models and methods for providing reliability of distributed
computing systems are expected to rest upon the notion of Quality of Service
(QoS) [4]. QoS is to describe the overall performance of a given network, in
particular its most important features, such as the data transfer rate (DTR),
the reliability of the network and its components, the IP packet transfer delay
(IPTD), the IP packet delay variation (IPDV), or jitter, and the IP packet loss
ratio (IPLR). Path reliability and IPTD, also called time in the system in this
paper, are of special interest for current research.
Although ways to provide reliability and fault-tolerance of computing net-
works have already been widely discussed, for example in [5–11], the opportu-
nities offered in this regard by dynamic distribution of data packets-requests
through the network, on the basis of multipath routing, remain understudied for
today.
This research paper discusses how to provide reliability of distributed com-
puting systems using redundant requests, spread through the network over mul-
tiple paths. It should be noted that current problem has nothing to do with
choosing the best route out of all possible ones; instead, it deals with utiliz-
ing numerous paths, not necessarily of the same cost, contemporaneously. Thus,
the study explores the methods of improving reliability by forwarding multi-
ple requests, or better to say multiple copies of the same request, along multiple
paths. It results, on the one hand, in the increased network load and hence in the
Redundant Distribution of Requests Through the Network 201
increased packet transfer delay, but on the other hand, in the enhanced chance
of a packet-request being delivered within a specified time interval.
There are proposed two ways of transmitting requests through a given net-
work by polling servers, which are possible receivers of the requests, in a serial
and in a simultaneous manner. These ways underlie the corresponding models,
which take into account the increased network load occurring because of simul-
taneous transmission of requests over several routes and the existence of the flow
of repeated (unserved) requests – the result of polling unavailable servers.
Fig. 1. The general scheme of searching for an available server. The router accepts the
Initial Request Flow and forwards incoming requests along one or a few appropriate
disjoint paths, with one switch (or router) on each way to the corresponding server.
Requests underlie the flow of either Accepted requests, which are to leave the sys-
tem after being processed by the server, or Rejected requests, turning into Repeated
requests when being distributed through the network again.
The proposed models take into account the possibility of loss of requests and
the probability p of server availability. Since server availability is not guaranteed,
there arises a problem of finding an available server. A server is considered to
be available and, as a result, ready to serve incoming requests if it is in working
order, powered up, and not overloaded with the requests having already come
202 V.A. Bogatyrev and S.A. Parshutina
from the network. While the arrived requests are being processed or waiting
on the server’s queue before entering service, the senders of those requests are
notified of success. If the server is busy or unavailable due to other reasons,
a request is rejected and its sender receives a notice of refusal, whereupon the
request is resent to another server over some different path. Servers are supposed
to be accessed repeatedly in a random or a Round-Robin fashion until the success
is achieved.
The general scheme of distribution of requests over n paths, with taking into
account the flow of rejected requests, which later resent to other servers, is shown
in Fig. 1. Requests forming the initial flow are forwarded along one or several
routes simultaneously. If a request is transmitted over one of all appropriate
paths at a time, it is serial search for available servers. Transferring a request over
two or more routes contemporaneously indicates parallel, or redundant, search.
On the way to one of n possible servers, each request passes through one switch
(or router) and gets into the flow of either accepted requests or rejected ones.
The accepted requests continue to the respective server and leave the system
after having been processed. The rejected requests underlie the flow of repeated
requests and subsequently merge with the flow of initial requests, newly arrived
in the system. The period of time a request resides in the system, including
the time spent on all unsuccessful efforts to be delivered and the successful
distribution over one of the routes in the end, is called requests’ time in the
system in this paper. It is chosen as the criterion of efficiency of the systems
under consideration.
The queueing model for distribution of requests through the network under
consideration is represented in Fig. 2. Here, Λ stands for the intensity of the ini-
tial flow of requests, which is divided into λ1 , λ2 , ..., λn flows; R1 , R2 , ..., Rn do
for switches (or routers) from 1 to n; P1 , P2 , ..., Pn are probabilities of servers’
S1 , S2 , ..., Sn being available for incoming requests. In this paper, the probabili-
ties are considered to be the same and equal to p.
Fig. 2. The queueing model of the system, where λ1 , λ2 , ..., λn are intensities of the
flows of distributed requests, λ1 P1 , λ2 P2 , ..., λn Pn are ones of accepted requests, and
λ1 (1 − P1 ), λ2 (1 − P2 ), ..., λn (1 − Pn ) are ones of rejected requests.
Time in the system, or average residence time, T (Λ) – replaced with (5) – is
∞
i−1 vtΛ + p[αvtΛ − n(v + t)]
T0 (Λ) = p i(1 − p) (T (Λ) + t) = . (6)
i=1
p[vΛ + p(αvΛ − n)]
Time delay arising from each effort to deliver a request, granting that Λ1 can
be replaced with (8) is
v vnr
T (Λ) = = . (9)
1− (kΛ1 (Λ)+αΛ)v r(n + αvΛ) − kvΛ
n
Fig. 3. Time in the system, or residence time of requests in the system, T, depending
on the intensity of request flow Λ, with server availability p = 0.95 (up) and p = 0.85
(down). Three ways of distributing requests through the network are illustrated, when
requests are transferred over one path at a time (solid line), over two paths (dashed
line) or over three paths (dash-dotted line) simultaneously.
in case p = 0.95 and Λ >= 0.5 1/s but is less attractive as compared with them
when p = 0.85 and Λ < 0.5 1/s.
It should also be noted that the redundant system with the redundancy order
k = 3 produces worse results than with k = 2 in case p = 0.95 but proves to
be better than the latter in case p = 0.85 and Λ < 0.25 1/s.
To conclude, the simultaneous use of several routes is pointless in case of
high server availability but, with its value decreasing, the efficiency of redundant
search starts to increase. However, this will continue until a certain point (the
threshold value), which is represented by the intersection of the functions of non-
redundant and redundant search in Fig. 3. On exceeding the threshold value,
the intensity of the flow of requests – of the initial ones as combined with the
repeated ones – all sent over multiple paths, causes the dramatic rise of time in
the system. Moreover, the lower server availability is, the greater the number of
those routes is, which are needed to detect an available server in the shortest
time.
Thus, an interesting problem arises, which is to optimize the choice of the
number of redundant paths, needed to distribute a request through the network,
with the use of multipath routing techniques. Solving this problem means making
use of the discussed strategies of transferring requests over k paths by applying
them in turn to non-uniform flows of requests.
4 Conclusion
In this paper, there are proposed models for fault-tolerant distributed computing
systems with adaptive distribution of requests through the network. The models
rest upon multipath routing techniques and describe the ways of polling servers:
– one at a time in case of serial (non-redundant) search and
– several servers at once in case of parallel (redundant) search.
A server is supposed to be available, i.e. ready for serving requests, or unavailable
due to its possible faults and temporary shutdown as well as overload with the
requests, which have already been accepted by the server and are either being
processed or waiting on its queue to be serviced.
Time in the system, or average residence time of requests in the system, takes
into account the flow of repeated requests (previously rejected by unavailable
servers) and is chosen as the criterion of efficiency of the systems in question.
It is found that non-redundant search for an available server, when each
request is sent over only one appropriate path at a time, while the others are
regarded to be reserve, turns out to be the best option if server availability is
high. The less server availability is, the more preferable redundant search is, when
each request is being transmitted over multiple paths simultaneously. However,
it is true only until the intensity of request flow is less than a certain threshold
value. Increasing the number of routes used to search for an available server (the
redundancy order) gives little advantage in case server availability is high but
proves significant if the availability is decreasing.
Redundant Distribution of Requests Through the Network 207
Acknowledgments. The study is a part of two research projects: “Methods and mod-
els for providing the integrated security and stable functioning of computer systems”
(State Registration Number is 610481) and “Methods for designing the key systems of
information infrastructure” (State Registration Number is 615869) of ITMO University,
St. Petersburg, Russia.
References
1. Gunnar, A., Johansson, M.: Robust load balancing under traffic uncertainty
tractable models and efficient algorithms. Telecommun. Syst. 48, 93–107 (2011)
2. Banner, R., Orda, A.: Multipath Routing Algorithms for Congestion Minimiza-
tion. CCIT Report No. 429, Dept. Electr. Eng., Technion, Haifa, Israel (2004).
http://www.ee.technion.ac.il/people/ron/Congestion.pdf
3. Hopps, C.: Analysis of an Equal-Cost Multipath Algorithm. IETF, RFC 2992
(2000). http://www.ietf.org/rfc/rfc2992
4. McDysan, D.: QoS and Traffic Management in IP and ATM Networks. McGraw-Hill,
New York (2000)
5. Andreev, S., Saffer, Z., Turlikov, A.: Delay analysis of wireless broadband net-
works with non real-time traffic. In: Sacchi, C., Bellalta, B., Vinel, A., Schlegel,
C., Granelli, F., Zhang, Y. (eds.) MACOM 2011. LNCS, vol. 6886, pp. 206–217.
Springer, Heidelberg (2011)
6. Bogatyrev, V.A.: Exchange of duplicated computing complexes in fault tolerant
systems. Autom. Control Comput. Sci. 46(5), 268–276 (2011)
7. Bogatyrev, V.A., Bogatyrev, S.V., Golubev, I.Y.: Optimization and the process of
task distribution between computer system clusters. Autom. Control Comput. Sci.
3, 103–111 (2012)
8. Bogatyrev, V.A.: Fault tolerance of clusters configurations with direct connection
of storage devices. Autom. Control Comput. Sci. 45(6), 330–337 (2011)
9. Bogatyrev, V.A.: An interval signal method of dynamic interrupt handling with
load balancing. Autom. Control Comput. Sci. 34(6), 51–57 (2000)
10. Bogatyrev, V.A.: Protocols for dynamic distribution of requests through a bus with
variable logic ring for reception authority transfer. Autom. Control Comput. Sci.
33(1), 57–63 (1999)
11. Bogatyrev, V.A., Bogatyrev, A.V.: Functional Reliability of a Real-Time Redun-
dant Computational Process in Cluster Architecture Systems. Autom. Control
Comput. Sci. 49(1), 46–56 (2015)
12. Kleinrock, L.: Queueing Systems, vol. 1: Theory. Wiley Interscience, New York
(1975)
13. Kleinrock, L.: Queueing Systems, vol. 2: Computer Applications. Wiley Inter-
science, New York (1976)
14. Vishnevsky, V.M.: Teoreticheskie osnovy proektirovaniya komputernykh setey
(Theoretical Fundamentals for Design of Computer Networks). Tekhnosfera,
Moscow (2003) (in Russian)
15. Aliev, T.I.: Osnovy modelirovaniya diskretnykh system (Fundamentals of simula-
tion of discrete systems). SPbSU ITMO Publ., St. Petersburg (2009) (in Russian)
FPGA-Prototyping with Advanced Fault
Injection Methodology for Tolerant Computing
Systems Simulation
1 Introduction
When designing fault-tolerant computing systems imposes high requirements
failure resiliency. In these systems, the failure of one element can lead to loss of
efficiency of the system. This defines one of the most important challenges fac-
ing the developers of such systems providing failure resiliency of the constituent
elements. This is especially important for embedded systems of unmanned space
vehicles, operating in conditions of space radiation. Therefore, the design and
testing steps of fault tolerant systems development have their own characteris-
tics. So these steps are carried out in three stages:
(1) development and functional testing of the system without failures protection
subsystems;
(2) development and implementation of failures protection subsystems and re-
conduct functional testing of system;
(3) checking operation of the system when adding faults and failures.
One way to perform this verification is a method of fault injection with
simulation of FPGA-based-prototyping of a computer system.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 208–223, 2016.
DOI: 10.1007/978-3-319-30843-2 22
FPGA-Prototyping with Advanced Fault Injection Methodology 209
Results analysis module. In this methodology, these steps do not have any special
features and is held normal job with the FPGA-based ASIC prototype.
Then we consider the functioning of the system in the third stage, i.e., using
the mechanism of fault injection.
Threats parameters are calculated on the basis of data about the effects of
technological parameters and the topology of the chip. Each threat is the list
of elements affected, time, duration of exposure and type of exposure (latching
values, setting a certain value, introducing delay).
Description of the test structures is a synthesized description of the test
structure that will allow on command from the control device to distort the
chip.
As discussed earlier in the modeling involved only the memory elements
(Fig. 3), so that all elements involved in the test should be replaced by a test
structure, examples of which are shown in figures (Figs. 4, 5 and 6). In the sim-
ulation using test patterns of the following types:
– fixed output value (Fig. 4) — after activating the OUTP DATA set to the
constant (0, 1 or a value stored in the element at the time of activation),
which persists until deactivating.
– a specified output value (Fig. 5) — after activation the OUTP DATA is set
to the value that was an input port TEST DATA. In the ongoing simulation
FPGA-Prototyping with Advanced Fault Injection Methodology 213
Fig. 4. Test structure with constant Fig. 5. Test structure with variable
error value error value
input value of TEST DATA, and accordingly the output OUTP DATA may
vary.
– the introduction of the delay (Fig. 6) — after activation of the test structure
passes a value from input DATA with a delay of 0.5–1 clock cycle input clock.
The modeling process uses two main models describing the internal and external
parameters of the chip, and providing information about possible threats to the
system.
214 O. Brekhov et al.
Fig. 9. Versions of the trajectories of heavy charged particles and damaged units
depends on the value of the largest amount of other elements is the most vul-
nerable, i.e. dangerous for the system.
Netlist dynamic analysis is carried out in the process which counts the num-
ber of items that changed their state by changing the state of the analyzed
element.
Developer guide suggest to create the list of elements, a failure in which, in the
opinion of the developer would lead to the most serious negative consequences.
Parallel Simulation of the Target VLSI Projects. In the first mode, all
expansion cards are used for independent parallel simulation of target VLSI
projects. The internal structures of the FPGAs for all cards will be the same
(Fig. 11): In addition to the target VLSI project the following modules are hard-
ware implemented:
– The Fault-Injection Module, that carries out direct fault injection to the VLSI
project accordingly to the fault injection pattern.
– Stimulus provider, imitating external signals for the target VLSI project
– The Data Acquisition Module, which reads the values of the outputs of the
target VLSI project and sends the data to the software-based simulation eval-
uation module for data processing
– The Control Unit, which coordinates all the modules implemented on FPGA,
starts and ends simulation and data exchange with the workstation.
RAM blocks, of the expansion cards allow to carry out temporary storage
of fault patterns, stimulus and output values of the chip. In general, this allows
to offload computing means of the workstation, and gives an opportunity to
perform long-term simulation without interruption.
In the parallel simulation mode it is possible to implement different scenarios:
Fig. 11. The internal structure of the FPGA in parallel mode simulation tolerant
architectures
Fig. 12. The internal structure of the master-FPGA in simulation systems, with TMR
by chip
Fig. 14. The internal design of the slave FPGA for huge ASIC prototyping
simulated systems will depend on the number of master FPGAs for each slave, as
well as the number of slaves.
Combined Simulation Mode. Combined simulation mode enables efficient
parallel simulation of VLSI projects of various sizes. For example, for VLSI
projects with a large number of elements the master - slave mode can be used,
and untapped FPGAs can be used to model projects of smaller volumes and
have an internal structure for the parallel simulation mode.
When simulating VLSI projects of large volumes the situation when a project
could not be implemented via the single FPGA may arise. In this case, by using
222 O. Brekhov et al.
Fig. 15. The internal design of the slave FPGA for huge ASIC prototyping for parallel
modeling
different partition algorithms the project composition is made to give the oppor-
tunity to implement the project via multiple FPGAs. In this case, depending on
the target VLSI project volume master-slave or parallel simulation modes could
be used. In the first case, the internal structure of the slave FPGA will have the
form shown in Fig. 14. In the case of parallel simulation scheme is also somewhat
more complicated (Fig. 15).
It is worth noting that the implementation of the calculations for models
of external influences and the emergence of threats due to the large volume of
data to be processed requires considerable computing power that could not be
implemented via separate HSSC. Solving this problem requires the development
of efficient algorithms of necessary calculations parallelization and the usage of
high-performance computing servers or networked systems of distributed com-
puting. The question of choosing the effective implementation of this approach
is not covered in this paper.
4 Conclusions
(1) In this paper, the method of modeling fault-tolerant computing systems based
on FPGA prototypes using the advanced fault-injection method is proposed.
(2) This method takes into account the parameters of the environment (types,
intensity and direction of radiation in the environment) and the parameters of
FPGA-Prototyping with Advanced Fault Injection Methodology 223
the internal environment (chip production technology and the topology of the
chip). Chip simulation, carried in the presence of specific injected threats shows
the behavior of the system under specified conditions.
(3) The concept of the integrated HSSC with defined simulation modes, the con-
stituent elements of the complex and a set of required software, including models
of the environment (external factors) and the occurrence of threats (appearance
of failures depending on internal factors) is proposed.
(4) The three-stage modeling concept, that combines functional testing and fail-
ures causes modeling is also proposed. It comprises:
1. functional test of the system without taking into account means of tolerating
faults and failures
2. functional test of the system with implement means of providing fault tolerance
3. verification of system performance in the presence of faults and failures.
References
1. Panasyuk, M.I., Novikov, L.S.: Space model. In: The Impact of the Space Environ-
ment on Spacecraft Materials and Equipment, vol. 2, Moscow (2007)
2. Tylka, A.J., Adams Jr., J.H., Boberg, P.R., Brownstein, B., Dietrich, W.F.,
Flueckiger, E.O., Petersen, E.L., Shea, M.A., Smart, D.F., Smith, E.C.: CREME96:
a revision of the cosmic ray effects on micro-electronics code. IEEE Trans. Nucl.
Sci. 44(6), 2150–2160 (1997)
3. Gaffey, J., Bilitza, D.: NASA/National space science data center trapped radiation
models. J. Spacecraft Rockets 31(2), 172 (1994)
4. Weller, R.A., Mendenhall, M.H., Reed, R.A., Schrimpf, R.D., Warren, K.M.,
Sierawski, B.D., Massengill, L.W.: Monte Carlo simulation of single event effects.
IEEE Trans. Nucl. Sci. 57(4), 1726–1746 (2010)
5. Mendenhall, M.H., Weller, R.A.: A probability-conserving cross-section biasing
mechanism for variance reduction in Monte Carlo particle transport calculations.
Nucl. Instrum. Meth. A 667(1), 38–43 (2012). doi:10.1016/j.nima.2011.11.084
6. Barak, J.: Simple calculations of proton SEU cross sections from heavy ion cross
sections. IEEE Trans. Nucl. Sci. 53(6), 3336–3342 (2006)
Space Simulation Wireless Broadband
Network in Electromagnetic Interference,
and Obstructions
Victor M. Churkov(B)
Famous figures CINR and RSSI measure the quality of the radio signal in the
client equipment. CINR (Carrier to Interference + Noise Ratio) - the ratio of
carrier signal level to noise level is used in telecommunications for assessing the
quality of the signal and connection.
If the signal quality is above a certain standard for the norm. RSSI (Received
Signal Strength Indication) - the power level of the received signal can be used
to approximate the signal quality estimate.
Indicator CINR measured periodically in the channel and radio signal receiver
is transmitted source for analysis and decision-making on how to connect and
uniquely defines the client’s network speed.
Thus, CINR is an integral indicator of the quality of the radio signal, taking
into account:
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 224–229, 2016.
DOI: 10.1007/978-3-319-30843-2 23
Space Simulation Wireless Broadband Network 225
its infrastructure, taking into account the relative position coordinate effectively
and to minimize the number of elements.
External factors - the desired signal level, noise, and electromagnetic distur-
bance largely determine the bandwidth of the radio channel, which can signif-
icantly reduce the data transmission rate and, consequently, the quality of the
traffic [1].
Spatial model of the wireless network comprises a matrix parameter CINR,
consider the quality of the radio as a function of given above influencing factors
defining network traffic at any point.
Formation of the matrix takes place based on the results of an experimental
study of the parameters of the radio signal CINR in the network coverage area.
Mathematical model [2] of the spatial organization of the coverage of wireless
broadband network comprises a plurality of functions, variables and constants
which: U - universal set of elements of the system wireless network (coverage):
S – A lot of system coverage subsets (regions);
C – The set of centers of areas;
R – Radius of the area;
(x, y, z) – Spatial coordinates;
u(x, y, z), c(x, y, z), s(x, y, z) – Elements of the set;
m, n – Cardinality of the set;
ρ(x, y, z) – An indicator of the effectiveness of the radio signal coverage.
Let U, C, S, R, (x, y, z), ρ(x, y, z) — variable mathematical model:
U, R, (x, y, z), ρ(x, y, z) – Asked variables;
C, S – Expanded variables.
Let the sets U, C, S associated functional dependency.
We introduce additional features f, φ, ψ, |νu | specifying constraints and per-
formance indicators for coverage, operating:
f – Computation C ⊂ U and subject to the limitations of the selected perfor-
mance criteria (for example |Sj | → max → |U |);
φ – Calculation of the elements subsets Ui ;
ψ – Calculation of the subset Uj \ Uj+1 ;
|−
ν→
u | – Calculation of magnitude of the vector in the coordinates U ;
Fig. 1. Distribution of SNR and network speed under nominal and operating noise.
Space Simulation Wireless Broadband Network 227
Fig. 3. The speed of the network when the standard 802.11n 5 GHz, standard 802.11ac
10 GHz.
5 Conclusion
The presented models and tools ensure the full, high-quality design and testing
of broadband wireless networks and systems for existing and future applications.
At the initial stage of the deployment of complex wireless networks, mod-
ernization and operation of these funds increase reliability; the stability of the
technical parameters, functional stability.
Reduces the complexity, cost, recovery time during operation and mainte-
nance.
References
1. Vishnevsky, V.M.: Broadband wireless data transmission networks, p. 592. Tech-
nosphera, Moscow (2005)
2. Churkov, V.M., Kosilov, N.A., Zhdanov, V.M.: Optimization of spatial models wire-
less sensor networks. In: Proceedings of Seventh International Conference Informa-
tion and Telecommunication Technologies Intelligent System, Schweiz, 03–09 July
2010, pp. 9–16 (2010)
Performance and Quality Evaluation of Cloud
Information Services
1 Introduction
1. Acknowledge the anomaly using NMC, SOD, or other detection methods and
sources listed in Table 1.
2. Communicate to the other team members and notify of the problem. For
that purpose, the RC Company uses the worldwide popular documenting
system JIRA [5] as well as in-house real time tracking web application Incident
Management Portal (IMP) with its own DB of incidents.
3. Restore the SaaS using predefined workaround in case the problem is well-
known, for example reboot problematic VM or redirect the workload to a
standby computing resource in the cloud.
4. Escalate the problem to corresponding technical support engineering if
IaaS/SaaS cant be restored by means of GNOC.
sent as well. Good solution for auto-remediation is described in [6] where capac-
ity management of Java applications, which are running under heavy workload
in server pool and often lead to a well-known problem of Java memory leak, is
introduced to prevent the entire SaaS outage and evaluate the required redun-
dancy of cloud computing resources.
select
ifnull(sum(tags_count)-count(*),0) as dupes_count
from
(select i.short_desc,
(select group_concat(t.tag_name)
from inc_has_tags h
left outer join service_tags t
on h.service_tags_id_tag = t.id_tag
where h.incident_inc_id = i.inc_id
group by h.incident_inc_id
) as tags,
count(*) as tags_count
from incident i
where create_start_date < curdate()
and create_start_date >= adddate(curdate(),-30)
group by i.short_desc, tags
having count(*) > 1
) as dupes_list;
Table 2 shows sample report generated in Zabbix with KPI metrics for IMP
and their daily values at RC.
Having KPI data collected in centralized DB separately for IMP/CMP/JIRA,
Zabbix then allows creating new integral metrics of arbitrary complexity, which
are computed locally on Zabbix side without additional remote querying.
Performance and Quality Evaluation of Cloud Information Services 233
system similar to IMP KPI described in Sect. 3. Table 3 shows sample report
generated in Zabbix with some CMP KPI metrics and daily values at RC.
Daily KPI reports are expanded to weekly, monthly, quarterly, etc. Zabbix
functionality allows defining the aggregated hosts/metrics and triggering the
alerts for specified thresholds in case overall degradation of IMP/CMP man-
agement process. Example in Fig. 2 shows monthly average statistics of CMR
implementation time. This CMP KPI metric is useful for analysis of the perfor-
mance and workload of daily planned maintenance. According to the graph, one
value has exceeded the 4 h limit that is not an incident and is allowed during
non-business weekend. On the other hand, the overall usage of daily maintenance
window is less than 50 % at RC Company and could be improved.
Paper [13] presents more examples of KPI evaluation specifically for big
dynamic enterprise environment with continuous delivery of SaaS updates to
the customers.
5 Conclusions
References
1. Mescheryakov, S., Shchemelinin, D., Efimov, V.: Adaptive control of cloud com-
puting resources in the internet telecommunication multiservice system. In: 6th
IEEE International Congress on Ultra Modern Telecommunications and Control
Systems, St. Petersburg (2014). http://ieeexplore.ieee.org/xpl/articleDetails.jsp?
arnumber=7002117
2. Zabbix Enterprise-Class Monitoring System. http://www.zabbix.com
3. RingCentral Cloud Services. http://www.ringcentral.com
Performance and Quality Evaluation of Cloud Information Services 237
4. Volkov, A., Efimov, V., Mescheryakov, S., Shchemelinin, D.: Integrated data model
for managing a multi-service dynamic infrastructure. In: Proceedings of the 3rd
International Conference on Computer Modeling and Simulation. Polytechnic Uni-
versity Publishing House, St. Petersburg (2014). http://dcn.icc.spbstu.ru/index.
php?id=344&L=2
5. Atlassian JIRA. https://www.atlassian.com/software/jira
6. Mescheryakov, S., Shchemelinin, D.: Capacity management of java-based business
applications running on virtualized environment. In: Proceedings of the 39th Inter-
national Conference for the Performance and Capacity by CMG, La Jolla, CA, USA
(2013). http://www.cmg.org/conference/cmg2013/
7. Fault Detection and Isolation. http://en.wikipedia.org/wiki/Fault detection and
isolation
8. Nonverbal Communication. http://en.wikipedia.org/wiki/Nonverbal communi
cation
9. Mean Time to Repair. http://en.wikipedia.org/wiki/Mean time to repair
10. Service Level Agreement. http://en.wikipedia.org/wiki/Service-level agreement
11. Performance Engineering. http://en.wikipedia.org/wiki/Performance engineering
12. Turnaround Time, Performance Management. http://en.wikipedia.org/wiki/
Turnaround time
13. Mescheryakov, S., Shchemelinin, D.: Big software deployments in a big enter-
prise environment. In: 2nd ASE International Conference on Big Data Science
and Computing, Stanford University, CA, USA (2014). http://www.ase360.org/
handle/123456789/135/
Heuristic Solution for the Optimal Thresholds
in a Controllable Multi-server Heterogeneous
Queueing System Without Preemption
1 Introduction
Controllable queueing systems with heterogeneous servers and single queue find
wide application in various fields of human activity including resource allocation
in telecommunication and computer networks, control problems in production
lines and so on. For detailed review of the literature on heterogeneous queues the
reader is referred to [2]. The queues with preemption or priority interruption of
D. Efrosinin—This work was funded by the Russian Foundation for Basic Research
(RFBR), Project No 15-08-08677-a.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 238–252, 2016.
DOI: 10.1007/978-3-319-30843-2 25
Heuristic Solution for the Optimal Thresholds 239
service, i.e. when the customer can change the server during the service process,
are well studied, see e.g. [6]. The optimal allocation mechanism for customers
in such systems is defined normally through a simple Fastest Free Server policy,
when a new customer is transfered to the fastest available server even if it is
already under service on some slower server. In this case the system can be
exhaustively studied independently of the number of servers in the system. In
contrast for the queueing systems without preemption the Fastest Free Server
allocation policy occurs to be not the best one. As it was shown in papers
[8,9] the optimal policy, which minimizes the long-run average cost per unit of
time, is of threshold type, i.e. for any server exists some threshold level, which
specifies the number of customers in the queue when the server must be activated.
There are a number of methods how to calculate the optimal threshold levels of
the control policy. They can be evaluated numerically by means of the Howard
iteration algorithm [5] or by numerical minimization of the average cost function
evaluated in closed form, see e.g. [3,4]. In both cases we have problems with
dimensionality of the model if the number of servers is relative high. Therefore
a natural question arises: Is there a possibility to get some heuristic solution
which will be valid for an arbitrary number of servers.
In framework of the paper we solve such a problem and construct formu-
las for the optimal threshold levels which seems to be quite appropriate for
the studied systems. The kernel element of the proposed heuristic method con-
sists in evaluation of the average cost function in closed form as a function of
unknown threshold levels. It is performed by the method of the difference equa-
tions. Then we analyze the boundaries between the optimality regions where the
fixed threshold level takes a certain value. It can be done at least for two and
three server models but obtained results can be generalized to the case of an
arbitrary number of servers.
The rest of the paper is organized as follows. Section 2 describes the math-
ematical model based on a homogeneous Markov process under fixed thresh-
old policy. In Sect. 3 the stationary state probabilities are evaluated by means
of the difference equations approach. Section 4 deals with optimization prob-
lem. Heuristic method for the optimal threshold level estimation is presented in
Sect. 5. Finally, some numerical examples are illustrated in Sect. 6.
In further sections we will use the notations ej for the vector with 1 in
the j-th (beginning from 0-th) position and 0 elsewhere, 1{A} for the indicator
function, where 1{A} = 1 if the condition A holds, and 0 otherwise, a, b for the
elements of the set [a, b] ∩ N0 .
2 Mathematical Model
Consider a queueing system where the customers arrive according to the Poisson
process with intensity λ and K heterogeneous servers have exponentially distrib-
uted service times with intensities μ1 , μ2 , . . . , μK . The service of the customers
is assumed to be without preemption, i.e. the customer being served on some
server can not change this server. The interarrival and service times are assumed
to be mutually independent. Define the control policy
240 D. Efrosinin and V. Rykov
f = (q1 , q2 , . . . , qK ),
which prescribes how to allocate the customers between the servers and is of
threshold type defined as a sequence of threshold levels
1 = q1 ≤ q2 ≤ · · · ≤ qK < ∞.
According to this policy the first k servers must be occupied whenever there are
q customers in the queue and q = qk , qk+1 − 1. The cost structure consists of the
following components:
c0 – the holding cost per unit of time for any waiting customer in the queue,
cj – the usage cost per unit of time for any busy server j.
The servers are enumerated in such a way that
0 < c1 μ−1 −1 −1
1 ≤ c2 μ2 ≤ · · · ≤ cK μK , (1)
0< μ−1
1 ≤ μ−1
2 ≤ ··· ≤ μ−1
K ,
Note that the state space of the servers ED (q) depends on the queue length q,
⎧ ⎫
⎨ dj ∈ {0, 1}, j = 1, K, q = 0 ⎬
ED (q) = d; dj = 1, di ∈ {0, 1}, 1 ≤ j ≤ k ≤ i − 1 ≤ K − 1, q = qk , qk+1 − 1 ,
⎩ ⎭
dj = 1, j = 1, K, q ≥ qK
Denote by Aqk (x) and Aqk (x), k = 1, K, the following events and their com-
plements,
Analyzing the transitions of the Markov process {X(t)}t≥0 and using the nota-
tions (4) we get the system of balance equations for the stationary state proba-
bilities πx , x = (q, d) ∈ E,
πx = lim P[X(t) = x]
t→∞
in the form
K
K
λ+ dk (x)μk πx = λ πx−e 1{A + πx−e 1
k qk −1 (x−ek )} 0 {Aq −1 (x−e0 )}
k=1 k=1 k
K
K
+ dk (x)μk πx+e 1{A
0 qk (x+e0 )}
+ (1 − dk (x))μk πx+e 1
k {Aqk (x+ek )}
. (5)
k=1 k=1
To get a solution of this system for the fixed threshold policy in closed form we
intend to apply the method of the difference equations.
π(0,0,0) = [Rq−1
2 +1
Bq2 +1 − 1]π(0,0,1) , (6)
π(q,1,0) = [Rk−q2 Bq2 +1 − Bq+1 ]π(0,0,1) , k = 0, q2 − 1,
π(q,1,1) = Bq+1 π(0,0,1) , q = 0, q2 − 1,
π(q,1,1) = Nq−q2 +1 π(q2 −1,1,1) , q ≥ q2
−1
Rq2 −1 Bq2 +1 N1 Bq2 +1
π(0,0,1) = Rq2 +1 Bq2 +1 + + ,
Rq2 +1 − Rq2 1 − N1
where
λ n λ n
Rn = , Nn = , Bn = b1 β1n + b2 β2n ,
μ1 μ1 + μ2
(λ + μ1 + μ2 ) ± (λ + μ1 + μ2 )2 − 4λμ1
β1,2 = ,
2μ1
1 − β1 β2 − 1
b1 = , b2 = .
β2 − β1 β2 − β1
Proof. Due to threshold structure of the control policy the system of balance
Eq. (5) are divided into separate subsystems which can be treated as homoge-
neous difference equations.
242 D. Efrosinin and V. Rykov
Subsystem 1.
The substitution of this solution to the equation of the previous system for the
arbitrary q and subsequent division by β q−1 lead to the equation
μ1 β 2 − (λ + μ1 + μ2 )β + λ = 0.
where
Bn = (b1 β1n + b2 β2n ).
From the first equation of the subsystem (6) for the probability π(0,1,1) we get
b1 β1 + b2 β2 = β1 + β2 − 1,
b1 + b2 = 1,
and substituting of this result to the equality (10), it is easy to show, that there is
exist the only one non-trivial solution of the quadratic equation, η = μ1 +μ λ
2
, i.e.
where
λ n
Nn = η n = .
μ1 + μ2
To solve the system the corresponding equations are summed up for the proba-
bilities π(q,1,0) and π(q,1,1) . As a result we obtain the following equations
π(q,1,0) = [Rq−q2 (b1 β1q2 +1 + b2 β2q2 +1 ) − (b1 β1q+1 + b2 β2q+1 )]π(0,0,1) (13)
= [Rq−q2 Bq2 +1 − Bq+1 ]π(0,0,1) , q = 0, q2 − 1.
Finally from
we get
μ1 q2 +1
π(0,0,0) = [ (b1 β1q2 +1 + b2 β2q2 +1 ) − 1]π(0,0,1) (14)
λ
= [Rq−1
2 +1
Bq2 +1 − 1]π(0,0,1) .
Theorem 2. For the three-server queueing system M/M/3 with a threshold pol-
icy f = (1, q2 , q3 ) the stationary state probabilities satisfy the following relations,
(R
q2 +1 − 1)Bq2 +1 (Nq3 −q2 +1 − N1 )Bq2
π(0,0,1,1) + π(0,0,1,0) = + (15)
Rq2 +2 − Rq2 +1 N1 − 1
Φ1 Hq3 Nq3 −q2 +1 Bq2 −1
+ ,
(1 − Φ1 )Hq3 +1
244 D. Efrosinin and V. Rykov
where
λ n λ n λ n
Rn = , Nn = , Φn = ,
μ1 μ1 + μ2 μ1 + μ2 + μ3
An = a1 α1n + a2 α2n , Bn = b1 β1n + b2 β2n , Gn = g1 γ1n + g2 γ2n , Hn = h1 ξ1n + h2 ξ2n ,
(λ + μ1 + μ2 + μ3 ) ± (λ + μ1 + μ2 + μ3 )2 − 4λμ1
α1,2 = ,
2μ1
(λ + μ1 + μ2 ) ± (λ + μ1 + μ2 )2 − 4λμ1
β1,2 = ,
2μ1
(λ + μ1 + μ3 ) ± (λ + μ1 + μ3 )2 − 4λμ1
γ1,2 = ,
2μ1
(λ + μ1 + μ2 + μ3 ) ± (λ + μ1 + μ2 + μ3 )2 − 4λ(μ1 + μ2 )
ξ1,2 = ,
2(μ1 + μ2 )
1 − α1 α2 − 1 1 − β1 β2 − 1
a1 = , a2 = , b1 = , b2 = ,
α2 − α1 α2 − β1 β2 − β1 β2 − β1
1 − γ1 γ2 − 1 L2 − L1
g1 = , g2 = , h1 = , h2 = 1 − h 1 ,
γ2 − γ1 γ2 − γ1 K1 − K2
1 1
K1 = − 1 ξ1q3 −q2 − − 1 ξ2q3 −q2 Aq2 ,
ξ2 ξ1
1
L1 = − 1 ξ2q3 −q2 Aq2 ,
ξ1
μ3 ξ1 − ξ1q3 −q2 +1 ξ2 − ξ2q3 −q2 +1 γ2q2 +1 − γ1q2 +1 1 1
K2 = Aq2 − − G−1 − ,
λ 1 − ξ1 1 − ξ2 q2
γ2 − γ1 ξ1 ξ2
μ3 −1 γ2q2 +1 − γ1q2 +1 1 μ3 ξ2 − ξ2q3 −q2 +1
L2 = G 1− Aq2 + R1−1 Aq2 +1 + Aq2 .
λ q2 γ2 − γ1 ξ1 λ 1 − ξ2
Heuristic Solution for the Optimal Thresholds 245
Proof. The proof can be performed in a similar way as before by dividing the
system of balance equations into subsystems, which are solved as homogeneous
difference equations.
where the notations q(x) and dj (x) stand for the elements of the vector state
x ∈ E. As it is known [11], for ergodic Markov process with costs the long-
run average cost per unit of time for the policy f coincides with corresponding
assemble average,
1
g f = lim V f (x, t) = c(y)πyf , (16)
t→∞ t
y∈E
where
t
V (x, t) =
f
Pf [X(u) = y|X(0) = x]c(y)du (17)
0 y∈E
denotes the total average cost up to time t when the process starts in state x
and πyf = Pf [X(t) = y] denotes a stationary probability of the process given
policy f . The policy f ∗ is said to be optimal when for any available policy f
∗
g f = min g f . (18)
f
if
λ
ρ = K < 1,
j=1 μj
which coincides with a stability condition of the system. Hence we can obtain
the explicit solution for the function g f := g(q2 , . . . , qK ) at least up to the case
of three servers. These formulas are used to get a heuristic solution.
246 D. Efrosinin and V. Rykov
Proof. Consider the case λ = 0. Due to our assumption about threshold structure
of the control policy f , we can calculate the total average cost until the system
becomes empty recursively. Assuming the known values of thresholds q2 , . . . , qk−1
we obtain the value of qk . Obviously for the state x = (0, . . . , 0) we have
V (x) = 0,
V (x + e1 ) = µc1 ,
1
V (x + e0 + e1 ) = c0µ+c1 + V (x + e1 ) = µc0 + 2c
µ1
1
,
1 1
...
(q −1)c +c q (q2 −1)c0 q2 c1
V (x + (q2 − 1)e0 + e1 ) = 2 µ 0 1 + V (x + (q2 − 2)e0 + e1 ) = 2 2µ + µ1
.
1 1
Heuristic Solution for the Optimal Thresholds 247
When the queue length has reached the level q2 , it becomes optimal to use the
second server. Then we have
c2 q2 (q2 − 1)c0 q2 c1
V (x + (q2 − 1)e0 + e1 + e2 ) = + + .
μ2 2μ1 μ1
Repeating the procedure up to the level qk one can obtain
k−1
k−1
(qk − qk−1 ) j=1 cj (qk − qk−1 )(qk−1 + qk − 1)c0
V x + (qk − 1)e0 + ej = k−1 + k−1 .
j=1 j=1 μj 2 j=1 μj
implies
k−1
ck (qk − qk−1 ) j=1 cj (qk − qk−1 )(qk−1 + qk − 1)c0
+ k−1 + k−1
μk j=1 μj 2 j=1 μj
k−1
(qk − qk−1 + 1) j=1 cj (qk − qk−1 + 1)(qk−1 + qk )c0
< k−1 + k−1 .
j=1 μj 2 j=1 μj
From the last inequality we get
1 ck
k−1 k−1
qk = μj − cj .
c0 μk j=1 j=1
1 ck 1 ck
k−1 k−1 k−1 k−1
μj − λ − cj ≤ q̂k∗ ≤ μj − cj . (21)
c0 μk j=1 j=1
c0 μk j=1 j=1
To prove the inequality at the right hand side it is necessary to show that
k−1
Fk ≤ μj .
j=1
To get the formulas for the case λ > 0 consider first the problem of the mean
number of customers minimization, i.e. when cj = 1, j = 0, K. To reduce the
number of system parameters in the system (5) we divide the right and the left
μ
hand side by λ and introduce the notations rj = λj , j = 1, K. The proposed
Algorithm 3 is based on a functional estimation of the boundaries between the
areas where the optimal threshold qk , k = 1, K, takes a certain value. Numerical
analysis confirms our expectations that the optimal threshold qk depends only
on parameters of the first k servers and the boundaries between the regions of
optimality have a linear structure, i.e. they can be represented as hyperplanes.
Algorithm 3.
Step 1. Calculation of the optimal thresholds qk , k = 1, K, for all possible values
of (r1 , . . . , rK ) by means of a function
K
N̄ := N̄ (r1 , . . . , rK , q2∗ , . . . , qK
∗
)= q(x) + dj (x) πx ,
x∈E j=1
k
akj rj + akk+1 = 0,
j=1
where the unknown coefficients are evaluated through coordinates of the two
points (x1 , y1 ) and (x2 , y2 ): a21 = y2 − y1 , a22 = x1 − x2 , a23 = x2 y1 − x1 y2 . By
selecting appropriate coordinates we get the coefficients summarized in the next
table:
Heuristic Solution for the Optimal Thresholds 249
i 1 2 3 4 5 6 7 8
x1 6.47 9.65 12.74 15.79 18.83 21.85 24.87 27.90
y1 3.00 3.00 3.00 3.00 3.00 3.00 3.00 3.00
x2 10.48 18.66 28.74 40.79 54.83 70.87 88.87 108.90
y2 5.00 6.00 7.00 8.00 9.00 10.00 11.00 12.00
a21 2.00 3.00 4.00 5.00 6.00 7.00 8.00 9.00
a22 −4.01 −9.01 −16.00 −25.00 −36.00 −49.01 −64.00 −81.00
a23 −0.91 −1.92 −2.96 −3.95 −4.98 −5.89 −6.96 −8.09
From the table one can see that the coefficients of the line can be represented
as a function of threshold estimation q̂2 ,
a21 (q̂2 ) = q̂2 + 1, a22 (q̂2 ) = −(q̂2 + 1)2 , a23 (q̂2 ) = −q̂2 .
To get the formulas for the model with costs we assume that it should have
the same structure as (19), where λ = 0. Hence we get the relation (19) for
λ > 0. In the same way one can derive the formulas for the case K = 3. For
250 D. Efrosinin and V. Rykov
three server model the regions of optimality for the levels q2 and q3 are the planes
as illustrated in Fig. 2. These planes are defined as
where the coefficients are expressed through the estimations q̂2 and q̂3 ,
a21 (q̂2 ) = q̂2 + 1, a22 (q̂2 ) = −(q̂2 + 1)2 , a23 (q̂2 ) = −q̂2 ,
a31 (q̂3 ) = a32 (q̂3 ) = q̂3 + 2, a33 (q̂3 ) = −(q̂3 + 2)2 , a34 (q̂3 ) = −q̂3 .
6 Numerical Examples
Example 1. Consider the queueing system M/M/5 arrival intensity λ = 0.9.
Other parameters take the following values:
j 0 1 2 3 4 5
cj 1.50 3.00 2.80 2.60 2.40 2.00
μj – 2.00 0.70 0.50 0.40 0.30
cj μ−1
j – 1.50 4.00 5.20 6.00 6.66
In the following table we list the results of calculations for the levels qk∗ and
q̂k∗ , k
= 2, K as well as the results for g in case of different control policies and
arrival intensity λ
Heuristic Solution for the Optimal Thresholds 251
The results show that the difference in performance between the OTP and HTP
does not exceed 1.5 % and these policies can be more than 25 % superior in
performance comparing to the mostly used in practice allocation policy F F S.
7 Conclusion
In this paper we have obtained heuristic relations for the optimal threshold levels
which minimize the long-run average cost per unit of time or, in particular,
the mean number of customers in the queue system with heterogeneous servers.
These formulas give satisfactory values for the levels qk , k = 2, K, with a relative
error which was less than 1.5 % in all realized examples. Also it was proved that
this policy is superior in performance comparing to some alternative allocation
policies. Therefore this policy can be treated at least as a quasi-optimal one. The
formulas were verified for different number of servers with quite equal relative
errors so we may expect the validity of these expressions to an arbitrary K.
These approach was successfully applied to the retrial heterogeneous queueing
systems as well, the results will be published shortly.
252 D. Efrosinin and V. Rykov
References
1. Aviv, Y., Federgruen, A.: The value-iteration method for countable state Markov
decision processes. Oper. Res. Lett. 24(5), 223–234 (1999)
2. Efrosinin, D.: Controlled Queueing Systems with Heterogeneous Servers. Dynamic
Optimization and Monotonicity Properties. VDM Verlag, Saarbrücken (2008)
3. Efrosinin, D.: Queueing model of a hybrid channel with faster link subject to partial
and complete failures. Ann. Oper. Res. 202(1), 75–102 (2013)
4. Efrosinin, D., Rykov, V.: On performance characteristics for queueing systems with
heterogeneous servers. Autom. Remote Control 69(1), 61–75 (2008)
5. Howard, R.: Dynamic Programming and Markov Processes. Wiley Series,
New York (1960)
6. Kumar, B.K., Arivudainambi, D.: Transient solution of an M/M/c queue with
heterogeneous servers and balking. Inf. Manage. Sci. 12(3), 15–27 (2001)
7. Puterman, M.L.: Markov Decision Process. Wiley Series in Probability and Math-
ematical Statistics. Wiley, New York (1994)
8. Rykov, V., Efrosinin, D.: Optimal control of queueing systems with heterogeneous
servers. Queueing Syst. 46, 389–407 (2004)
9. Rykov, V., Efrosinin, D.: On the slow server problem. Autom. Remote Control
70(12), 2013–2023 (2009)
10. Sennott, L.I.: Stochastic Dynamic Programming and the Control of Queueing Sys-
tems. Wiley, New York (1999)
11. Tijms, H.C.: Stochastic Models. An Algorithmic Approach. Wiley, New York
(1994)
Evaluation of Functionality’s Efficiency
in Priority Telecommunication Networks
with Heterogeneous Traffic
1 Introduction
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 253–259, 2016.
DOI: 10.1007/978-3-319-30843-2 26
254 I. Kalinin and L. Muravyeva-Vitkovskaya
data packets, speech data packets, text data packets and etc.), where differ-
ent delivery requirements exist [3]. This circumstance must be kept in network
administrator’s mind for increasing using TCN’s resources efficiency. Traffic’s
prioritizing is one of the ways to distribute network resources according to exist-
ing priorities.
Building networks using switches allows enabling network technology inde-
pendent traffic’s prioritization for increasing user traffic’s servicing quality. This
new capability in comparison with networks built only upon hubs is consequence
of buffering data frames in switches before sending them to another port. Usually
switch keeps not one but some queues for every input and output ports, where
every queue has it’s own processing priority.
Queue processing algorithms are the one of the basic methods of QoS in
network elements.
Solving TCN design problems assumes using efficient models and mathemat-
ical methods, which give an opportunity to make qualitative and quantitative
analysis of TCN’s functioning characteristics depending on structural, functional
and load parameters. In the TCN’s analyzing process one of the major charac-
teristics being determined is delivery time from sender to recipient. Results,
represented in [4], give on opportunity to calculate only average delivery time
for various message types. But in practice not average time but probability of
prompt delivery from sender to recipient for different message types is more
interesting (e.g. operative, service, dialog, file message types). Moreover, it’s
necessary to keep in mind a possibility of using priority techniques for informa-
tion flows management, where priority techniques are based on general discipline
with mixed priorities.
To solve a given task an open queue network with heterogeneous request
flow can be used. Let’s illustrate method of analyzing and getting relatively sim-
ple results assuming nodes having non-priority information flows management
methods.
2 Method
Problem Definition. Let messages of K types circulate in TCN with n nodes.
Let us assume that abonent, connected to sender node h, generates markov-
ian exponential flow of k -type messages with λk (0 |h, l) intensity to destination
recipient node l, this kind of messages will be called “(h, l )-messages of k -type”
for short. Also let us assume that probability πk (i, j |h, l) of transferring (h, l )-
messages of k -type from node i to connected node j is defined based on chosen
routing algorithm. For each (h, l )-direction these probabilities πk (i, j |h, l) form
transfer probability matrix which describes possible transfer routes from node h
to node l (i, j, h, l = 1, n; k = 1, K).
Message processing duration in node i, represents processing time in node
and transferring time to neighbor node, will be assumed exponentially dis-
tributed with central tendency equals to bk (i). Different type messages are
Priority Telecommunication Networks with heterogeneous Traffic 255
[1 − R(i)][1 + s bk (i)]−1 s
Uk∗ (i, s) = K , (5)
s − Λ(i) + r=1 λr (i)/[1 + s br (i)]
K K
where Λ(i) = k=1 λk (i); R(i) = k=1 λk (i) bk (i); i = 1, n; k = 1, K; s > 0.
The Laplace transformation Vk∗ (h, l, s) for probability density of delivery time
of k-type messages from node h to node l defines by the following set of equations:
n
Vk∗ (h, l, s) = Uk∗ (h, s) πk (h, j|h, l) Vk∗ (j, l, s) (h, l = 1, n), (6)
j=1
(2)
where Uk (h) and Uk (h) — respectively first and second initial moments of stay
duration of k-type messages in node h. These moments are defined by derivation
(5) by s at point s=0.
3 Results
It’s possible to determine various probabilistic and pertaining to time telecom-
munication system’s characteristics, specifically, probability of message’s prompt
delivery [6], using Laplace transformations Vk∗ (h, l, s) or moments Vk (h, l) and
(2)
Vk (h, l). Prompt delivery probability equals value of delivery time Laplace
transformation calculated with s = sk if message’s aging function is exponential
and average aging time for k-type messages equals 1/s.
For (i, j|h, l) = (1, 2|1, 2); (1, 3|1, 3); (3, 4|1, 4) (2, 3|2, 3); (3, 1|2, 1);
(3, 4|2, 4); (2, 1|3, 1); (2, 4|3, 4); (3, 2|3, 2); (2, 1|4, 1); (3, 4|2, 4); (3, 1|4, 1);
Example. Let’s examine TCN, containing four nodes and two types of messages
circulating in it. Markovian exponential flow of k-type messages generates with
λk (0|h, l) intensity from abonents, connected to sender node h, to destination
recipient node l.
0, 02 if h < l 0, 03 if h < l
λ1 (0|h, l) = ; λ (0|h, l) = ; h, l = 1, 4.
0, 04 if h > l 2 0, 01 if h > l
Processing durations are the same and equals 2 s. for all messages in every
node. Delivery times for k-type messages from sender node h to destination
node l are constrained by V̂k (h, l) = 50 s for all k = 1, 2; h, l = 1, 4, where
h = l. Average aging time for k-type message equals 10 s (k = 1, 2). Probabili-
ties πk (i, j|h, l), shown in Table 1, are defined based upon chosen routing algo-
rithms where πk (0, h|h, l) = πk (l, 0|h, l) = 1 (i, j = 1, 4; h, l = 1, 4; k = 1, 2).
For this TCN we have message’s prompt delivery probabilities P (Vk (h, l) <
V̂k (h, l)) (h, l = 1, 4; k = 1, 2), shown in Table 2, and prompt delivery probabili-
ties, shown in Table 3, granting aging functioning.
TCN’s characteristic’s computation method, based on decomposition, also
can be used in case of priority message management methods used in TCN’s
nodes. Stay durations for nodes can be determined as described in [7] in case of
using mixed priorities message management methods. Final results, computed
by this method, are approximate, because message flows of different types at
output, and thus at input, differ from exponential in case of priority manage-
ment methods. However result’s inaccuracies lie in acceptable for engineering
computations limits as was discovered in research of flow’s characters and their
influencing on results in wide range of parameters corresponding to real systems.
258 I. Kalinin and L. Muravyeva-Vitkovskaya
K H l
1 2 3 4
1 1 1,000 0,996 0,989 0,984
2 0,994 1,000 0,973 0,994
3 0,985 0,973 1,000 0,985
4 0,973 0,996 0,989 1,000
2 1 1,000 0,996 0,989 0,704
2 0,764 1,000 0,073 0,764
3 0,749 0,973 1,000 0,749
4 0,712 0,996 0,989 1,000
K h L
1 2 3 4
1 1 0,667 0,374 0,327 0,233
2 0,354 0,561 0,275 0,354
3 0,312 0,275 0,490 0,312
4 0,211 0,374 0,327 0,667
2 1 0,667 0,374 0,327 0,109
2 0,174 0,561 0,275 0,174
3 0,158 0,275 0,490 0,158
4 0,113 0,374 0,327 0,667
4 Conclusions
Found results can be used for solving TCN’s optimization problem, which lies
in routing algorithm determination (transfer probabilities πk (i, j|h, l)) and in
assigning priorities to messages of different types, providing specified message
delivery time.
Described method of TCN’s computation is implemented in program system.
References
1. Jakubovich, D.: Network traffic optimization. Commun. Netw. Syst. 10, 92–97
(2001)
2. Aliev, T.I.: Characteristics of mixed priorities message servicing disciplines. Izv.
vuzov. Priborostroenie 4(57), 30–35 (2014)
3. Goldstein B.S., Pinchuk A.V., Suhovickiy, A.L. VoIP. – M.: Radio and communica-
tion (2001)
Priority Telecommunication Networks with heterogeneous Traffic 259
4. Basharin, G.P., Tolmachev, A.L.: Queue network’s theory and its application to
analysis of information-calculating systems. Science and Engineering Summaries.
Probability Theory. Mathematical Statistics. Theoretical Cybernetics. – M.: VINITI
T. 21, p. 3–119 (1983)
5. Kleinrock, L.: Computation systems with queues: Russian translation. – M.: Mir
(1979)
6. Zaharov, G.P.: Research methods of data networks. – M.: Radio and communication
(1982)
7. Aliev, T.I., Muravyeva-Vitkovskaya, L.A.: Prioritetnye strategii upravleniya
trafikom v multiservisnykh komp’yuternykh setyakh [Priority-based strategies of
traffic management in multiservice computer networks]. Izv. vuzov. Priborostroenie
54(6), 44–48 (2011)
Estimation Quality Parameters of Transferring
Image and Voice Data over ZigBee
in Transparent Mode
1 Introduction
Today, Ubiquitous Sensor Networks (USN) are deeply integrated in the everyday
life of an ordinary person. Research in the field of capability, network security,
traffic characteristics are actively conducted by large number of scientists in
different countries of the world [1–3].
Currently, in the interaction of the FUSN, within the concept of the Internet
of Things [4,5], there are a number of unresolved issues. One of them is the
delivery of various types of traffic based on the USN.
Recently a new direction of development in USN is Flying Ubiquitous Sensor
Networks (FUSN) [6] has become popular. The scope of FUSN is beneficial for
industry, agriculture, transport and monitoring of the functioning over various
objects.
FUSN is a part of USN, thereby it has a similar principles of organization
and architecture. Additionally, this networks have peer-to-peer or hierarchical
(cluster) models. Also, there are 2 segments of FUSN: terrestrial and flying.
Being a part of the FUSN USN, it has a similar organization principles and
architecture. Such networks can be peer or hierarchical (cluster). Also allocate
2 segments FUSN: ground and flying, which in turn can be peer or hierarchical.
FUSNs are realized by ZigBee specification, 6 LoWPAN, RPL, Bluetooth
Low Energy (BLE) and so on [7,8], which enables self-organizing networks with
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 260–267, 2016.
DOI: 10.1007/978-3-319-30843-2 27
Estimation Quality Parameters of Transferring Image 261
clusters topologies, and low cost realization, transferring small amounts of infor-
mation and which is characterized as a low power consumption.
The self-organizition in Flying Ubiquitous Sensor Networks (FUSN) is an
automatic creation of network topology, self-acting connection of new nodes,
automatic choice of routes the packet network without human intervention and
this is the one of typical sides of ZigBee specification. The main feature of FUSN
is a large coverage of sensors fields for terrestrial segment. Therefore, there is a
possibility of clustering of the terrestrial segment and the using of UAVs as a
main node.
The public unmanned aerial vehicles (UAV-P) can be used for FUSN creation
and FUSN-P is the network abbreviation in this case.
The voice and video transmission from terrestrial segment to UAV-P is the
next investigation task because often it may be the only chance to pass the
necessary information to the area of terrestrial sensor fields. There is a positive
experience of transfer of voice data over the protocol ZigBee [9]. In connection
with this problem the comparison of voice quality and video quality of experience
during their transmission from the terrestrial segment for years using different
protocols (ZigBee, 6LoWPAN, RPL), and different languages is of great interest
for FUSN-P.
2 Goal of Investigation
To assess the quality of voice in different languages and videos from the terrestrial
segment, we have chosen the ZigBee protocol, whereas it is commonly known,
has many applications in various fields. It is planned to conduct researches and
compare with the others protocols.
ZigBee networks are created on the base of nodes of 3 types: coordinators,
routers and end points [10]. Coordinator is a generating network, forming and
functioning as the control center and network trust center - setting security policy
defining the settings in the process of joining devices to the network, responsi-
ble for security keys. Router is transferring packets of data, realizing dynamic
routing, restoring routes on network congestion or failure of any device. Routers
are connected to the coordinator or others routers on the forming of network,
and can be attached to child devices — routers or end points. Router works in
continuous mode, has permanent power consumption and can provide up to 32
end devices. End point can send and receive packets, but does not translate and
route. End points can connect to the coordinator or router, but cannot have child
devices. Additionally, end points can be converted into sleep mode to conserve
battery power. Developers of this specification have documented of transferring
small packets of data, mostly text and packets insensitive for delays.
Despite on this, the goal was to analyze the opportunities of transferring
multimedia data over ZigBee with appropriate quality. Investigation possibilities
of transmitting data such as voice, video, image will expand the range of services
to end users on the basis of WPAN networks.
262 R. Kirichek et al.
3 Experimental Evaluation
To achieve the goal, the following tasks were allocated: investigation of existing
algorithms and approaches to transferring data via FUSN; development labora-
tory test bed of transferring image data and voice; the experiment of broadcast
image and voice data; evaluation of the obtained results quality over the method
of assessing the quality of perception, according ITU recommendations [11,12].
It is known, that developed solutions for voice channel through ZigBee have
already existed and studied, but the commercial development which is performed
for specific hardware and software technology are not universal, and the solutions
do not involve an assessment of transmission quality, according to the existing
standards.
For practical implementation and experimentation laboratory test bed was
assembled based on debugging kits by company Silicon Labs, which was based
on ZigBee modules by Telegesis. Two devices of ETRX3 were selected, satis-
fying the research conditions, that enable managing the network through the
AT command and quickly establishing a connection via asynchronous receiver-
transmitter (UART). Since the task is to translate the voice and the image, test
beds have to include different components, besides the main part of research —
the ZigBee network. The wireless network was installed between two comput-
ers, which have simulated a transparent channel (Transparent mode). Scheme of
laboratory test bed shown in Fig. 1.
The initial data was selected as an image size of 37 KB, it is obvious that
the resolution of the image isn’t large, but it is more than enough to recognize
objects and analyze their actions. After event from software information is input
to the ZigBee ETRX3. The image then is transmitted to the coordinator via the
wireless network. The next step is transferring bytes to the computer through a
UART interface that can operate at a speed of 115200 bits per second, using the
hardware flow control. ETRX3 modules can transmit data in two modes. The
first approach is “AT+DMODE” which simulates transparent mode working.
This method is transmitted data without acknowledgment, thus a higher data
rate is 12.7 Kbit per second. The second method allows transferring numbered
packets and confirms the correct data sending. It helps to recover lost packets;
this method is “AT+SCAST”. Thereby, for analyze network possibilities we
needed to take into account different parameters, such as indicators of quality
perception, delays and losses, thus a second mode have selected for laboratory
test bed —“AT+SCAST”. If the loss had occurred, the resulting image would
have damaged and as a result, would have not displayed. “AT+SCAST” allows
broadcasting data from the router to the coordinator, which is marked in the
network as a Sink. This module is part of the network ZigBee, which performs
a function of receiving all transmitted data over the network. “AT+SCAST”
provides correct transferring and receiving image without distortion, but the
speed of this process is much slower than the first mode. The speed was 7.5 Kbit
per second. The original image was transmitted in 39 s.The next experiment has
a direction on voice over ZigBee network; the scheme of laboratory test bed is
shown on Fig. 2.
The goal of this experiment was to transmit voice in real time. The parame-
ters of the experiment are the same as in the previous case, but this test bed has
included a microphone for inputting information and a speaker for outputting
information. Each experiment was performed at 15 s, the experts were speak-
ing into a microphone in different languages, and the voice is transmitted to an
analog-digital converter and then is converted into the PCM format. A method
for transmitting data via network has selected “AT+DMODE”, because this
type of traffic is not sensitive to losses, and is more dependent on the speed
of transmission and delays. The sampling rate was 8000 Hz with 8 quantization
bits, and was sending the data to the transmission buffer every 10 ms. Therefore,
one packet contained the 80 bits. These characteristics ensured low sound qual-
ity; moreover collection on the receiving side for full audio output was failed.
Thereby, test bed was using voice compression codec. Lossy compression algo-
rithm A-Law was chosen, compliant with G.711 [13]. Due to the compression of
traffic was able to increase quantization up to 16-bit for filling each packet of 80
bits. After transferring data to the PCM format at the transmitting side of test
bed traffic, it was compressed and then was passed via transparent channel. The
receiver side is a decompressed data and output to the speaker, where Assistant
assessed results. One of the important cases, connecting with realization net-
work solutions, is provisioning quality of service for each service. Additionally,
the requirements for network transmission parameters are specific to different
applications and types of traffic. Recommendation ITU-T Y.1540 [14] has iden-
tified the following characteristics of the network as the most important in terms
of their impact on the quality of service: throughput, network reliability, delays
and delay variability, loss ratio, network survivability asses.
This research has analyzed the influence of network characteristics on the
quality of voice and image transmission over a wireless sensor network, according
to the existing recommendations. This will determine the further expediency of
264 R. Kirichek et al.
using the ZigBee specification for the transmission of multimedia traffic in similar
technologies.
The most interesting thing is the transfer of the image, because today most
services are focused on the visualization of information. But the widespread usage
of Ubiquitous Sensor Networks suggests that image transmission over ZigBee
becomes ordinary technology. For example, if there is a sensor movement in the
house or in a protected area, you will be interested to get the picture of what or
who caused the alarm. If it is the hare or other animal, there is no necessity to
call special services.
Obviously, the existing Sensor Networks are problematic to transmit image
or video with acceptable levels of quality of service. But for the implementation
of applications such as “smart infrastructure”, the values of some parameters
will be enough.
For the quality assessment of the image and voice transmission a methodology
was chosen proposed in the ITU-T Recommendation R.913. As a subjective eval-
uation method, method of Absolute Category Rating (ACR) was chosen. This
method uses a categorical estimation. The test sequence is evaluated according
to the established scale of assessments. The advantage of the method is the abil-
ity to evaluate the ACR which only receives test sequence on the receive side
without the etalons, so it gets you closer to the real conditions of the network
and the estimated of end-users.
The experiment evaluated such parameters as transmission speed, the num-
ber of losses, the quality of speech recognition in multiple languages. Evaluations
were conducted by four experts. The all duration was 15 s. The five experiments
for each language were provided. The decision to perform experiments on the
transmission of speech using fragments in different languages was made on the
basis of the fact that the quality of speech in each language demands different
requirements. Subjective assessment may vary from one language to another, due
to the fact that languages differ in sound, some are more melodic, some are more
clear and easy to recognize and understand, etc. That is why the experiments
were conducted: Russian, English, French, Arabic and Belarusian languages. To
reduce the chance of exhibiting lower valuation of the expert to foreign speech,
at least one native speaker of each language is included in the expert group. The
expert estimates are determined according to the following five-point scale: 5 —
excellent; 4 — good; 3 — acceptable; 2 — bad; 1 — unacceptable.
The first stage of the results analysis was calculation of the average rating
for each demonstration by the formula (1).
N
1
ujkr = uijkr (1)
N i=1
Similarly, the confidence intervals were calculated for the other languages.
The results are combined in Table 2.
Evaluation of subjective methods performed to research the connection
between objective indicators of the network and the subjective perception of
information by users based on changes during transmission. For displaying the
results the logistic curve approximation based on the function (4) was used. The
result is shown in Fig. 3.
5
σ= (4)
1 + e −t−t
B
0
The next step in the analysis was to determine the relationship between
the results. To determine the correlation of subjective assessments transmission
quality, lossy and speed of transmission were calculated by the correlation coef-
ficient. Based on the fact, that the coefficient of correlation between subjective
assessment and loss - 0.9, and between the speed of transmission - 0.3, it can be
concluded that the estimates are more dependent on the number of lost packets,
rather than on the speed of transmission over the network ZigBee.
4 Conclusions
– In case of low speed and high losses, quality of voice transmission is signifi-
cantly inferior of traditional communication networks. The number of losses
provides a large impact of transferring data via ZigBee.
– At that moment transferring video over ZigBee is problematic in the case of
low bandwidth of channel. The transmission of low resolution images makes
rational requirements for latency and jitter.
– On the one hand, the compression of voice by codecs with losses of significant
information increases probability of packet delivery. On the other hand, if a
packet is lost in the network, user will lose a significant number of data. The
choice of codec with enhanced compression algorithm will improve the results.
Thus, the scope of the using ZigBee networks for transmitting multimedia
information is specific. These networks are useful where requirements to the
quality of voice and image transmission rate are low. The main advantage of
using a networking standard ZigBee is a low cost, high autonomy, simple creation
and survivability of these networks, which allows us to consider the networks for
Estimation Quality Parameters of Transferring Image 267
the transmission of multimedia data in the future. Video and voice are the only
one way to communicate over FUSN in some cases. Especially, if it is realized in
the countryside or outback. Whereas quality of transferring voice and image is
satisfactory for all tested languages, this direction can take important place in
FUSN.
References
1. Koucheryavy, A., Prokopiev, A.: Ubiquitous sensor networks traffic models for
telemetry applications. In: Balandin, S., Koucheryavy, Y., Hu, H. (eds.) NEW2AN
2011 and ruSMART 2011. LNCS, vol. 6869, pp. 287–294. Springer, Heidelberg
(2011)
2. Vybornova, A., Koucheryavy, A.: Ubiquitous sensor networks traffic models for
medical and tracking applications. In: Andreev, S., Balandin, S., Koucheryavy,
Y. (eds.) NEW2AN/ruSMART 2012. LNCS, vol. 7469, pp. 338–346. Springer,
Heidelberg (2012)
3. ITU-T Recommendation Y.2060: Overview of Internet of Things. Geneva, February
2012
4. Lera, A., Floerkemeier, C., Mitsugi, J., Morabito, G.: The internet of things. IEEE
Wirel. Commun. 17(6), 44–51 (2010)
5. Kirichek, R., Paramonov, A., Koucheryavy, A.: Flying ubiquitous sensor networks
as a quening system. In: Proceedings of International Conference on Advanced
Communication Technology (ICACT 2015), Phoenix Park, Korea (2015)
6. IEEE Standard for Local and metropolitan area networks–Part 15.4: Low-Rate
Wireless Personal Area Networks (LR-WPANs):IEEE 802.15.4 (2011)
7. RPL: IPv6 Routing Protocol for Low-Power and Lossy Networks: RFC 6550– 03.
2012
8. Touloupis, A., Meliones, S.: Apostolacos speech codes for high-quality voice over
ZigBee applications: evaluation and implementation challenges. IEEE Commun.
Mag. 50(4), 122–128 (2012)
9. ZigBee Alliance, ZigBee Specification, 17 January, 2008
10. ITU-R Recommendation BT.500-13: Methodology for the Subjective Assessment
of the Quality of Television Pictures. Geneva, January 2012
11. ITU-T Recommendation P.910: Subjective video quality assessment methods for
multimedia applications. Geneva, April 2008
12. ITU-T Recommendation G.711: Pulse code modulation (PCM) of voice frequen-
cies. Amendment 1: New Annex A on lossless encoding of PCM frames (1998)
Amd.1 (08/2009)
13. ITU-T. Recommendation Y.1540: Internet protocol data communication service–
IP packet transfer and availability performance parameters. Geneva, November
2007
14. ITU-T Recommendation P.913: Methods for the subjective assessment of video
quality, audio quality and audiovisual quality of Internet video and distribution
quality television in any environment. Geneva, January 2014
Low-Priority Queue Fluctuations in Tandem
of Queuing Systems Under Cyclic Control
with Prolongations
1 Introduction
Conflicting traffic flows control at a crossroad is one of classical problems in queu-
ing theory. In the literature several algorithms were investigated: fixed duration
cyclic algorithm, cyclic algorithm with a loop, cyclic algorithm with changing
regimes, etc [1–6]. However, several (two in our case) consecutive crossroads are
of great interest, because in a real-life situation a vehicle having passed one high-
way intersection finds itself at another one. In other words, an output flow from
the first intersection forms an input flow of the second intersection. Hence, the
second input flow no longer has an a priori known simple probabilistic structure
(for example, that of a non-ordinary Poison flow), and knowledge about the ser-
vice algorithm should be taken into account to deduce formation conditions of
the first output flow.
Tandems of intersections were considered by a few authors. In [7] a computer-
aided simulation of adjacent intersection was carried out. In [8] a mathematical
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 268–279, 2016.
DOI: 10.1007/978-3-319-30843-2 28
Low-Priority Queue Fluctuations in Tandem of Queuing Systems 269
We assume that fj (z) converges for any z ∈ C such that |z| < (1 + ε), ε > 0.
(j)
Here pν is the probability of a bulk size in flow Πj being exactly ν = 0, 1, . . . .
Having been serviced the customers from O1 come back to the system as the Π4
customers. The Π4 customers in turn after service enter the system as the Π2
ones. The flows Π2 and Π3 are conflicting in the sense that their customers can’t
be serviced simultaneously. This implies that the problem can’t be reduced to a
problem with fewer input flows by merging the flows together.
In order to describe the server behavior we fix positive integers d, n0 , n1 , . . ., nd
and we introduce a finite set Γ = {Γ (k,r) : k = 0, 1, . . . , d; r = 1, 2, . . . nk } of states
server can reside in. At the state Γ (k,r) sever stays during constant time T (k,r) .
Define disjoint subsets Γ I , Γ II , Γ III , and Γ IV of Γ as follows. In the state γ ∈ Γ I
only customers from the queues O1 , O2 and O4 are serviced. In the state γ ∈
Γ II only customers from the queues O2 and O4 are serviced. In the state γ ∈ Γ III
only customers from queues O1 , O3 , and O4 are serviced. In the state γ ∈ Γ IV only
customers from queues O3 and O4 are serviced. We assume that Γ = Γ I ∪ Γ II ∪
Γ III ∪ Γ IV . Set also 1 Γ = Γ I ∪ Γ III , 2 Γ = Γ I ∪ Γ II , 3 Γ = Γ III ∪ Γ IV .
The server changes its state according to the following rules. We call a set
Ck = {Γ (k,r) : r = 1, 2, . . . nk } the k-th cycle, k = 1, 2, . . ., d. For k = 0 the state
Γ (0,r) with r = 1, 2, . . ., n0 is called a prolongation state. Put r ⊕k 1 = r + 1
for r < nk , and r ⊕k 1 = 1 for r = nk (k = 0, 1, . . ., d). In the cycle Ck we
select a subset CkI of input states, a subset CkO of output states, and a subset
CkN = Ck \ (CkO ∪ CkI ) of neutral states. After the state Γ (k,r) ∈ Ck \ CkO the
server switches to the state Γ (k,r⊕k 1) within the same cycle Ck . After the state
Γ (k,r) in CkO the server switches to the state Γ (k,r⊕k 1) if number of customers
270 V. Kocheganov and A.V. Zorine
⎧
⎪
⎪
⎪ Γ (k,r⊕k 1) if Γ (k,r) ∈ Ck \ CkO or (Γ (k,r) ∈ CkO ) ∧ (y > L);
⎪
⎨Γ (0,h1 (Γ (k,r) )) if Γ (k,r) ∈ CkO and y L;
h(Γ (k,r) , y) = (2)
⎪Γ (0,r⊕0 1)
⎪ if k = 0 and y L;
⎪
⎪
⎩h (r) if k = 0 and y > L.
3
Fig. 2. A tandem of crossroads, the physical interpretation of the queuing system under
study
3 Mathematical Model
The queuing system under investigation can be regarded as a cybernetic control
system what helps to rigorously construct a formal stochastic model [8]. The
scheme of the control system is shown in Fig. 1. There are following blocks present
in the scheme: (1) the external environment with one state; (2) input poles of
the first type — the input flows Π1 , Π2 , Π3 , and Π4 ; (3) input poles of the
second type — the saturation flows Π1sat , Π2sat , Π3sat , and Π4sat ; (4) an external
memory — the queues O1 , O2 , O3 , and O4 ; (5) an information processing device
for the external memory — the queue discipline units δ1 , δ2 , δ3 , and δ4 ; (6) an
internal memory — the server (OY); (7) an information processing device for
internal memory — the graph of server state transitions; (8) output poles —
the output flows Π1out , Π2out , Π3out , and Π4out . The coordinate of a block is its
number on the scheme.
Let us introduce the following variables and elements along with their value
ranges. To fix a discrete time scale consider the epochs τ0 = 0, τ1 , τ2 , . . . when
272 V. Kocheganov and A.V. Zorine
the server changes its state. Let Γi ∈ Γ be the server state during the interval
(τi−1 ; τi ], κj,i ∈ Z+ be the number of customers in the queue Oj at the instant
τi , ηj,i ∈ Z+ be the number of customers arrived into the queue Oj from the
flow Πj during the interval (τi ; τi+1 ], ξj,i ∈ Z+ be the number of customers in
the saturation flow Πjsat during the interval (τi ; τi+1 ], ξ j,i ∈ Z+ be the actual
number of serviced customers from the queue Oj during the interval (τi ; τi+1 ],
j ∈ {1, 2, 3, 4}.
The server changes its state according to the following rule:
where the mapping h(·, ·) is defined by formula (2). To determine the duration
Ti+1 of the next time slot it useful to introduce a mapping hT (·, ·) by
A functional relation
between ξ j,i and κj,i , ηj,i , ξj,i describes the service strategy. Further, since
We also have from the problem settings the following relations for the flow Π4 :
η4,i = min{ξ1,i , κ1,i + η1,i }, κ4,i+1 = κ4,i + η4,i − η2,i , ξ4,i = κ4,i . (6)
Put κi = (κ1,i , κ2,i , κ3,i , κ4,i ). The non-local description of the input and
saturation flows consists of specifying particular features of the conditional
probability distribution of selected discrete components ηi = (η1,i , η2,i , η3,i , η4,i )
and ξi = (ξ1,i , ξ2,i , ξ3,i , ξ4,i ) of marked point processes {(τi , νi , ηi ); i 0} and
{(τi , νi , ξi ); i 0} with marks νi = (Γi ; κi ). Let ϕ1 (·, ·) and ϕ3 (·, ·) be defined
by series expansions
∞
z ν ϕj (ν, t) = exp{λj t(fj (z) − 1)}
ν=0
with functions fj (z) defined by (1), j ∈ {1, 3}. The function ϕj (ν, t) equals the
probability of ν = 0, 1, . . . arrivals in the flow Πj during time t 0. If ν < 0
the value of ϕj (ν, t) is set to zero. Define function ψ(·, ·, ·) by
Then ψ(k; y, pk,r ) is the probability of k arrival from flow Π2 given the queue
O4 contains y customers and the server state is Γ (k,r) . For values k ∈ {0, 1, . . . , y}
the value of ψ(k; y, u) is set to zero.
Let a = (a1 , a2 , a3 , a4 ) ∈ Z4+ and x = (x1 , x2 , x3 , x4 ) ∈ Z4+ . If the mark value
is νi = (Γ (k,r) ; x) then the probability ϕ(a, k, r, x) of simultaneous equalities
η1,i = a1 , η2,i = a2 , η3,i = a3 , η4,i = a4 according the problem statement is
ϕ1 (a1 , hT (Γ (k,r) , x3 ))·ψ(a2 , x4 , pk̃,r̃ )·ϕ3 (a3 , hT (Γ (k,r) , x3 ))·δa4 ,min {(k̃,r̃,1),x1 +a1 }
Theorem 2. Let Γ0 (ω) = Γ (k,r) ∈ Γ and κ3,0 (ω) = x3,0 ∈ Z+ be fixed. Then
the stochastic sequence (7) is a homogeneous denumerable Markov chain.
P({ω : Γi+1 (ω) = Γ (k̃,r̃) , κ3,i+1 (ω) = x̃} | {ω : Γi (ω) = Γ (k,r) , κ3,i (ω) = x})
= (1 − δx̃3 ,0 ) · ϕ x̃3 + (k̃, r̃, 3) − x3 , hT Γ (k,r) , x3
(k̃,r̃,3)−x3
+ δx̃3 ,0 ϕ3 a, hT Γ (k,r) , x3 .
a=0
The next theorem clarifies which states of the Markov chain {(Γi , κ3,i ); i 0}
are essential. To make a complete classification we introduce sets
nk
3
S0,r = (Γ (0,r) , x3 ) : x3 ∈ Z+ , x3 > L − max (k, t, 3) , 1 r n0 ,
k=1,2,...,d
t=0
r
3
Sk,r = (Γ (k,r) , x3 ) : x3 ∈ Z+ , x3 > L − (k, t, 3) , 1 k d, 1 r nk .
t=0
x̃3 +(k̃,r̃,3)
Q3,i+1 (γ̃, x̃3 ) = (1 − δx̃3 ,0 ) × Q3,i (γ, x3 ) × ϕ3 (x̃3 + (k̃, r̃, 3)
x3 =0 γ∈H−1 (γ̃,x3 )
(k̃,r̃,3)
(k̃,r̃,3)−x3
− x3 , T (k̃,r̃) ) + δx̃3 ,0 Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
(8)
Suppose k and r are such that Γ (k,r) ∈ Γ . Let’s define partial probability
generating functions
∞
M (k, r, v) =
(i)
Q3,i (Γ (k,r) , w)v w ,
w=0
∞
qk,r (v) = v −(k,r,3) ϕ3 (w, T (k,r) )v w .
w=0
(k̃,r̃,3)
(k̃,r̃,3)−x3
α̃i (k̃, r̃, v) = Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
(k̃,r̃,3)
(k̃,r̃,3)−x3
− Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=0
L
αi (0, r̃, v) = α̃i (0, r̃, v) + q0,r̃ (v) × Q3,i (Γ (k1 ,r1 ) , x3 )
x3 =0
L
+ Q3,i (Γ (0,r0 1) , x3 ) v x3 − qk̃,r̃ (v) Q3,i (Γ (0,r0 1) , x3 ), Γ (0,r̃) ∈ Γ.
x3 =0
L
αi (k̃, r̃, v) = α̃i (k̃, r̃, v) − qk̃,r̃ (v) Q3,i (Γ (k,rk 1) , x3 )
x3 =0
+ Q3,i (Γ (0,r2 )
, x3 ) v x3 , Γ (k̃,r̃) ∈ Ck̃I
L
αi (k̃, r̃, v) = α̃i (k̃, r̃, v) − qk̃,r̃ (v) Q3,i (Γ (k,rk 1) , x3 )v x3 , Γ (k̃,r̃) ∈ Ck̃O ∪ Ck̃N
x3 =0
Theorem 5. Following recurrent w.r.t. i 0 relations take place for the partial
probability generating functions:
1. Γ (0,r̃) ∈ Γ , r̃ = 1, n0
2. Γ (k̃,r̃) ∈ Γ , k̃ = 1, d, r̃ = 1, nk̃
M(i+1) (k̃, r̃, v) = qk̃,r̃ (v) × M(i) (k̃, r̃ k̃ 1, v) + αi (k̃, r̃, v);
(k̃,r̃,3)
(k̃,r̃,3)−x3
= Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
∞
w+(k̃,r̃,3)
+ Q3,i (γ, x3 ) × ϕ3 (w + (k̃, r̃, 3) − x3 , T (k̃,r̃) )v w .
w=1 x3 =0 γ∈H−1 (γ̃,x3 )
(9)
After changing summation order by x3 and w second summand splits into two
∞ w+( k̃,r̃,3)
Q3,i (γ, x3 ) × ϕ3 (w + (k̃, r̃, 3) − x3 , T (k̃,r̃) )v w
w=1 x3 =0 γ∈H−1 (γ̃,x3 )
(k̃,r̃,3) ∞
= Q3,i (γ, x3 ) × ϕ3 (w + (k̃, r̃, 3) − x3 , T (k̃,r̃) )v w
x3 =0 w=1 γ∈H−1 (γ̃,x3 )
∞
∞
+ Q3,i (γ, x3 ) × ϕ3 (w + (k̃, r̃, 3)
x3 =(k̃,r̃,3)+1 w=x3 −(k̃,r̃,3) γ∈H−1 (γ̃,x3 )
− x3 , T (k̃,r̃) )v w . (10)
(k̃,r̃,3) ∞
Q3,i (γ, x3 ) × ϕ3 (w + (k̃, r̃, 3) − x3 , T (k̃,r̃) )v w
x3 =0 w=1 γ∈H−1 (γ̃,x3 )
(k̃,r̃,3)
∞
= Q3,i (γ, x3 ) ϕ3 (w + (k̃, r̃, 3) − x3 , T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=1
(k̃,r̃,3)
∞
= Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w .
x3 =0 γ∈H−1 (γ̃,x3 ) w=(k̃,r̃,3)+1−x3
(11)
Low-Priority Queue Fluctuations in Tandem of Queuing Systems 277
Using the same calculations for the second summand one can derive
∞
∞
Q3,i (γ, x3 )
x3 =(k̃,r̃,3)+1 w=x3 −(k̃,r̃,3) γ∈H−1 (γ̃,x3 )
(k̃,r̃,3)
∞
− Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w (12)
x3 =0 γ∈H−1 (γ̃,x3 ) w=0
Substitute (11) and (12) to (10) and after that (10) to (9), we get:
(k̃,r̃,3)
(k̃,r̃,3)−x3
M(i+1) (k̃, r̃, v) = Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
(k̃,r̃,3)
∞
+ Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=(k̃,r̃,3)+1−x3
∞
∞
+ Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=0
(k̃,r̃,3)
∞
− Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w .
x3 =0 γ∈H−1 (γ̃,x3 ) w=0
(k̃,r̃,3)
(k̃,r̃,3)−x3
M(i+1) (k̃, r̃, v) = Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
(k̃,r̃,3)
∞
+ Q3,i (γ, x3 )v x3 −(k̃,r̃,3) [ ϕ3 (w, T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=(k̃,r̃,3)+1−x3
∞
∞
− ϕ3 (w, T (k̃,r̃) )v w ] + qk̃,r̃ (v) Q3,i (γ, x3 )v x3
w=0 x3 =0 γ∈H−1 (γ̃,x3 )
and consequently
278 V. Kocheganov and A.V. Zorine
(k̃,r̃,3)
(k̃,r̃,3)−x3
M(i+1) (k̃, r̃, v) = Q3,i (γ, x3 ) ϕ3 (a, T (k̃,r̃) )
x3 =0 γ∈H−1 (γ̃,x3 ) a=0
(k̃,r̃,3)
(k̃,r̃,3)−x3
− Q3,i (γ, x3 )v x3 −(k̃,r̃,3) ϕ3 (w, T (k̃,r̃) )v w
x3 =0 γ∈H−1 (γ̃,x3 ) w=0
∞
+ qk̃,r̃ (v) Q3,i (γ, x3 )v x3
x3 =0 γ∈H−1 (γ̃,x3 )
∞
= α̃i (k̃, r̃, v) + qk̃,r̃ (v) Q3,i (γ, x3 )v x3 . (13)
x3 =0 γ∈H−1 (γ̃,x3 )
∞
Consider sum x3 =0 γ∈H−1 (γ̃,x3 ) Q3,i (γ, x3 )v x3 in more details depending
on γ̃. In case γ̃ = Γ (0,r̃) from definition H−1 (·, ·) one can derive H−1 (γ̃, x3 ) =
{Γ (k1 ,r1 ) , Γ (0,r̃0 1) } for x3 L and H−1 (γ̃, x3 ) = ∅ for x3 > L. Here pair (k1 , r1 )
such that h1 (Γ (k1 ,r1 ) ) = r̃. The sum is transformed to
∞ L
Q3,i (γ, x3 )v x3 = Q3,i (Γ (k1 ,r1 ) , x3 ) + Q3,i (Γ (0,r0 1) , x3 ) v x3 .
x3 =0 γ∈H−1 (γ̃,x3 ) x3 =0
(14)
In case γ̃ ∈ Ck̃I from definition H−1 (·, ·) one finds that H−1 (γ̃, x3 ) = ∅ for
x3 L and H−1 (γ̃, x3 ) = {Γ (k̃,r̃k̃ 1) , Γ (0,r2 ) } for x3 > L. Here r2 such that
h3 (r2 ) = Γ (k̃,r̃) . The sum is transformed to
∞
∞
Q3,i (γ, x3 )v x3 = Q3,i (Γ (k,rk 1) , x3 ) + Q3,i (Γ (0,r2 ) , x3 ) v x3
x3 =0 γ∈H−1 (γ̃,x3 ) x3 =L+1
∞
= Q3,i (Γ (k,rk 1) , x3 ) + Q3,i (Γ (0,r2 ) , x3 ) v x3
x3 =0
L
− Q3,i (Γ (k,rk 1) , x3 ) + Q3,i (Γ (0,r2 ) , x3 ) v x3 .
x3 =0
(15)
In the last case γ̃ ∈ Ck̃O ∪ Ck̃N , one can derive H−1 (γ̃, x3 ) = {Γ (k̃,r̃k̃ 1) } for
all x3 0. The sum takes following form
∞
∞
Q3,i (γ, x3 )v x3 = Q3,i (Γ (k,rk 1) , x3 )v x3
x3 =0 γ∈H−1 (γ̃,x3 ) x3 =L+1
∞
L
= Q3,i (Γ (k,rk 1) , x3 )v x3 − Q3,i (Γ (k,rk 1) , x3 )v x3 .
x3 =0 x3 =0
(16)
Substituting (14), (15) and (16) to the (13) one gets theorem statement.
Low-Priority Queue Fluctuations in Tandem of Queuing Systems 279
Acknowledgments. This work was fulfilled as a part of State Budget Research and
Development program No. 01201456585 “Mathematical modeling and analysis of sto-
chastic evolutionary systems and decision processes” of N.I. Lobachevsky State Uni-
versity of Nizhny Novgorod and was supported by State Program “Promoting the
competitiveness among world’s leading research and educational centers”
References
1. Neimark, Y., Fedotkin, M.A., Preobrazhenskaja, A.M.: Operation of an automate
with feedback controlling traffic at an intersection. Izvestija of USSR Academy of
Sciences, Technical Cybernetic (5), 129–141 (1968)
2. Fedotkin, M.A.: On a class of stable algorithms for control of conflicting flows or
arriving airplanes. Prob. Control Inf. Theor. 6(1), 13–22 (1977)
3. Fedotkin, M.A.: Construction of a model and investigation of nonlinear algorithms
for control of intense conflict flows in a system with variable structure of servicing
demands. Lith. Math. J. 17(1), 129–137 (1977)
4. Litvak, N.V., Fedotkin, M.A.: A probabilistic model for the adaptive control of
conflict flows. Autom. Remote Control 61(5), 777–784 (2000)
5. Proidakova, E.V., Fedotkin, M.A.: Control of output flows in the system with cyclic
servicing and readjustments. Autom. Remote Control 69(6), 993–1002 (2008)
6. Afanasyeva, L.G., Bulinskaya, E.V.: Mathematical models of transport systems
based on queueing theory. Trudy of Moscow Institute of Physocs and Technology
(4), 6–21 (2010)
7. Yamada, K., Lam, T.N.: Simulation analysis of two adjacent traffic signals.
In: Proceedings of the 17th Winter Simulation Conference, pp. 454–464. ACM,
New York (1985)
8. Zorin, A.V.: Stability of a tandem of queueing systems with Bernoulli noninstanta-
neous transfer of customers. Theor. Probab. Math. Stat. 84, 173–188 (2012)
9. Kocheganov, V.M., Zorine, A.V.: Probabilistic model of tandem of queuing systems
under cyclic control with prolongations. In: Proceedings of Internation Conference
Probability Theory, Stochastic Processes, Mathematical Statistics and Applications,
Minsk, 23–26 Feburary 2015, pp. 94–99 (2015)
On Conditional Monte Carlo Estimation of Busy
Period Probabilities in Gaussian Queues
1 Introduction
The self-similar nature of broadband traffic [7] has a deep impact in terms of
network dimensioning and Quality of Service (QoS) issues [3]. Indeed, persistence
phenomena (known in the literature as Noah effect) imply that the arrival rate
can remain on relatively high values for a considerable amount of time. Such
a behavior, closely related to the duration of busy periods in the underlying
queueing system, negatively affects QoS performance in terms of loss probability
and distribution of losses.
In traffic modelling, Gaussian processes have emerged as a flexible and pow-
erful tool, able to take into account the long memory properties of real traffic,
This work is supported by Russian Foundation for Basic research, projects 15–07–
02341 A, 15–07–02354 A,15–07–02360 A, and also by the Program of strategic devel-
opment of Petrozavodsk State University.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 280–288, 2016.
DOI: 10.1007/978-3-319-30843-2 29
On Conditional Monte Carlo Estimation of Busy Period Probabilities 281
50
40
30
Xt
20
10
0
0 10 20 30 40 50
t
Fig. 1. The sample path of the process X lies above the diagonal
while keeping a relatively simple and elegant description. The best known model
is Fractional Brownian Motion (FBM), originally proposed by Norros [10], but
our method is more general and includes FBM as a special case.
In more detail, in this work we consider a centered Gaussian process with
stationary increments {Xt , t ∈ R+ }. Let us denote by vt := VarXt the variance
of Xt ; then the covariance function has the following expression:
1
Γs,t = vt + vs − v|t−s| . (1)
2
We are interested in the estimation of the following probability
π(T) := P (∀t ∈ T : Xt > t), (2)
where T = [0, T ] ⊆ R+ (see Fig. 1). Such probability is closely related to the
duration of busy periods and plays an important role in the study of QoS indexes
since it takes into account bursts of losses, see [9,11] for more details.
It is worth mentioning that our approach only requires that vt is an increasing
function. Such condition is quite general and holds for the following processes,
well-known in the literature:
1. Fractional Brownian Motion (FBM). In this case vt = t2H , with Hurst para-
meter H ∈ (0, 1); in the teletraffic framework usually H ∈ (0.5, 1), corre-
sponding to traffic processes with long-range dependence. It has been shown
in [13] that FBM arises as the scaled limit process when the cumulative work-
load is a superposition of on-off sources with mutually independent heavy-
tailed on and/or off periods.
2. Sum of independent FBMs with vt = i t2Hi . The use of this model is also
motivated by the fundamental result in [13] in case of heterogeneous on-off
sources.
282 O. Lukashenko et al.
The key contribution of this work is the application of a variant of the condi-
tional Monte Carlo method for variance reduction, to estimate the target proba-
bility (2) when T → ∞. Indeed in this case the event {∀t ∈ T : Xt > t} becomes
rare and hence standard Monte Carlo requires an unacceptable large number of
generated sample paths.
Note that the analytical expression of the target probability is not known in
explicit form for general Gaussian input (including considered examples). Indeed,
there are only a few asymptotic results available based on the large deviation
theory. For instance in case of FBM input
1
lim log P (∀t ∈ (0, T ] : Xt > t)
T →∞ T 2−2H
1 Xt
= lim log P ∀t ∈ (0, 1] : √ > t = − inf I(f ), (3)
n→∞ n n f ∈B
where
B := {f ∈ R : f (r) > r, ∀r ∈ (0, 1]},
I denotes the rate function and R is the reproducing kernel Hilbert space
(RKHS) associated with the distribution of FBM (see [2] for more details).
π = P(X ∈ A) = EI(X ∈ A)
for some Borel set A of the paths of the process X, where I denotes the indicator
function. To estimate π by standard Monte Carlo simulation, we should generate
N replications X1 , ..., XN of the process X and calculate the sample mean
1
N
π
N = In .
N n=1
1
N
πN = E[Z|Yn ]. (4)
N n=1
Note that the variance of this estimator is always less than the variance of
the standard Monte Carlo one since
VarZ = E[Var[Z|Y ]] + Var[E[Z|Y ]]. (5)
3 BMC Estimator
The Bridge Monte Carlo (BMC) is a special case of the conditional Monte Carlo
method, particularly suitable for the estimation of the rare event probabilities
in a queueing system with Gaussian input.
Originally proposed by some of the authors in [4,5], BMC is based on the
idea of expressing the overflow probability as the expectation of a function of the
Bridge Y := {Yt } of the Gaussian input process X, i.e., the process obtained by
conditioning X to reach a certain level at some prefixed (deterministic) time τ :
Yt = Xt − ψt Xτ , (6)
where function ψt is expressed via covariance function as
Γt,τ
ψt := .
Γτ,τ
Because the variance of the input is an increasing function of t in all models
we consider, it is easy to see that ψt > 0 for all t ∈ T . Moreover, we note that
the process Y is independent of Xτ since
Γt,τ
E[Xτ Yt ] = Γτ,t − Γτ,τ = 0.
Γτ,τ
The BMC approach, originally proposed for the estimation of the overflow
probability [8], can also be adopted for estimating our target probability π(T).
Indeed,
π(T) := P (∀t ∈ T : Xt > t)
= P (∀t ∈ T : Yt + ψt Xτ > t)
t − Yt
= P ∀t ∈ T : Xτ >
ψt
t − Yt
= P Xτ ≥ sup = P Xτ ≥ Y ,
t∈T ψt
284 O. Lukashenko et al.
Recall that
Xt =d N (0, Γt, t ) =d Γt, t N (0, 1),
where =d stands for stochastic equivalence. Then the considered probability can
be rewritten as follows
π = P Xτ ≥ Y = P(Xτ ≥ u)P(Y ∈ du)
R
u
= P N (0, 1) ≥ P(Y ∈ du)
R Γτ, τ
Y
=E Φ ,
Γτ, τ
where independence between Y and Xτ is used and Φ denotes the tail distribu-
tion of standard normal variable, that is
∞
1 2
Φ(x) = √ e−y /2
dy.
2π x
(i)
Hence, given an i.i.d. sequence {Y , i = 1, ..., N } distributed as Y , the
estimator of π(T), by (9), is defined as follows:
(i)
1 N
Y
πNBMC
= Φ . (8)
N i=1 Γτ,τ
Note that
Y
Φ = E I(Xτ > Y )|Y , (9)
Γτ, τ
therefore the BMC approach is actually a special case of the conditional Monte
Carlo method. By (5), VarZ ≥ Var[E[Z|Y ]], so we can expect that the BMC
estimator implies variance reduction (comparing to the Crude Monte Carlo sim-
ulation) in the estimation of the target probability π(T).
4 Simulation Results
In this section, we point out through simulation results the accuracy of the BMC
estimator and the dependence of its performance on different parameters. For
sake of brevity, we only present results for FBM input. Moreover, in the following
we consider N = 10000 replications and a discrete lattice T = {0, 1, ..., T }.
On Conditional Monte Carlo Estimation of Busy Period Probabilities 285
1e−07
Probability (log scale)
1e−09
1e−11
Figure 2 shows the dependence of the target probability on the time duration
T in case of H = 0.8. The probability π(T ) exhibits an exponential decay in
agreement with the known large deviations asymptotic results, see formula (3).
Figure 3 highlight that π(T ) strongly depends on the Hurst parameter H. It is
quite natural that in the short-range dependent case (H < 1/2) the probability
the process X is above the diagonal for some fixed time duration T (in this
set of simulation T = 10) is extremely small, while in presence of long-range
dependence (that is H ∈ (1/2, 1)) π(T) is much higher.
286 O. Lukashenko et al.
0.18
0.16
Relative Error
0.14
0.12
0.10
In order to verify the goodness of our estimator, Figs. 4 and 5 show the
dependence of the relative error on parameters T and H respectively. Although
the relative error is not bounded (indeed, it is even unknown whether BMC is
asymptotically efficient), it grows slowly, and for the target probabilities of the
order of 10−11 is still less than 18 % (compare for example the values in Figs. 2
and 4).
Finally, Fig. 6 highlights the dependence of the relative error on the con-
ditioning time τ , which was used for constructing the Bridge process Y .
On Conditional Monte Carlo Estimation of Busy Period Probabilities 287
0.40
0.30
Relative Error
0.20
0.10
Empirical results highlight that it is better to choose larger values of free para-
meter τ in order to reduce the relative error.
5 Conclusions
In this paper, we addressed the busy period duration in Gaussian queue and esti-
mated the related probabilities by means of the BMC approach, a special case
of conditional Monte Carlo estimator. The BMC simulation approach exploits
the Gaussian nature of the input process (independence is equivalent to uncor-
relatedness) of the input process and relies on the properties of bridges.
Several simulation experiments have been carried out in order to study the
properties of the proposed estimator in case of FBM traffic for different values of
the relevant parameters (duration of the interval, value of the Hurst parameter
and choice of the conditioning point).
As a further research step, it is possible to study the influence of so-called dis-
cretization step on simulation results and the asymptotic efficiency/inefficiency
of BMC estimator.
References
1. Addie, R., Mannersalo, P., Norros, I.: Most probable paths and performance formu-
lae for buffers with Gaussian input traffic. Eur. Trans. Telecommun. 13, 183–196
(2002)
2. Deuschel, J.D., Stroock, D.W.: Large Deviations. Academic Press, San Diego
(1989)
288 O. Lukashenko et al.
3. Erramilli, A., Narayan, O., Willinger, W.: Experimental queueing analysis with
long-range dependent packet traffic. IEEE/ACM Trans. Netw. 4(2), 209–223 (1996)
4. Giordano, S., Gubinelli, M., Pagano, M.: Bridge Monte-Carlo: a novel approach
to rare events of Gaussian processes. In: Proceedings of the 5th St. Petersburg
Workshop on Simulation, pp. 281–286, St. Petersburg, Russia (2005)
5. Giordano, S., Gubinelli, M., Pagano, M.: Rare events of gaussian processes: a per-
formance comparison between Bridge Monte-Carlo and importance sampling. In:
Koucheryavy, Y., Harju, J., Sayenko, A. (eds.) NEW2AN 2007. LNCS, vol. 4712,
pp. 269–280. Springer, Heidelberg (2007)
6. Kulkarni, V., Rolski, T.: Fluid model driven by an Ornstein-Uhlenbeck process.
Probab. Eng. Info. Sci. 8, 403–417 (1994)
7. Leland, W.E., Taqqu, M.S., Willinger, W., Wilson, D.V.: On the self-similar nature
of Ethernet traffic (extended version). IEEE/ACM Trans. Netw. 2(1), 1–15 (1994)
8. Lukashenko, O.V., Morozov, E.V., Pagano, M.: Performance analysis of Bridge
Monte-Carlo estimator. Trans. KarRC RAS 3, 54–60 (2012)
9. Mandjes, M., Norros, I., Glynn, P.: On convergence to stationarity of fractional
Brownian storage. Ann. Appl. Probab. 19, 1385–1403 (2009)
10. Norros, I.: On the use of fractional Brownian motion in the theory of connectionless
networks. IEEE J. Sel. Areas Commun. 13(6), 953–962 (1995)
11. Norros, I.: Busy periods for fractional Brownian storage: a large deviation app-
roach. Adv. in Perf. Anal. 2, 1–19 (1999)
12. Ross, S.M.: Simulation. Elsevier, Amsterdam (2006)
13. Taqqu, M.S., Willinger, W., Sherman, R.: Proof of a fundamental result in self-
similar traffic modeling. Comput. Commun. Rev. 27, 5–23 (1997)
Stability and Admissible Densities
in Transportation Flow Models
1 Introduction
... < zN (t) < ... < z1 (t) < z0 (t) (1)
For this infinite chain of cars we try to find control mechanism which garanties
that the distance between any pair of neighbouring cars is greater or equal (on
all time interval (0, ∞)) to some fixed number d, called safe distance, but at
the same time is not too far from it (then we say that the density of cars is
admissible).
This control mechanism is assumed to be local - any car has information only
about the previous car. Moreover, this mechanism is of physical nature, like forces
between molecules in crystals but our “forces” are not symmetric. The safety
(stability) conditions appear to be similar to the dynamical phase transition in
the model of the molecular chain rapture under the action of external force.
However here we do not need the double scaling limit, used in [9].
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 289–295, 2016.
DOI: 10.1007/978-3-319-30843-2 30
290 A.A. Lykov et al.
We would like to organize such traffic so that for any t and k the distances
rk (t) = zk−1 (t) − zk (t) were greater or equal to some number d > 0, which is
chosen to avoid collisions and keep maximal possible density of traffic. Moreover,
the organization should use only (maximally) local control. More exactly, the k-
th driver at any time t knows only its own coordinate and velocity and the
coordinate zk−1 (t) of the previous car. Thus, for any k ≥ 1 the trajectory zk (t)
is uniquely defined by the trajectory zk−1 (t) of the previous particle.
Using physical terminology one could say that if, for example, rk (t) becomes
larger than d, then some positive force Fk increases acceleration of the particle
k. We will see that, besides Fk , for such stability, also friction force −αvk (t),
which, on the contrary, restrains the growth of the velocity vk (t), is necessary.
The constant α > 0 should be chosen appropriately.
Thus the trajectories are uniquely defined by the system of equations for
k≥1
dzk dzk
zk (t) = Fk (t) − α = ω 2 (zk−1 (t) − zk (t) − d) − α (3)
dt dt
where Fk is taken to be simplest possible
Stability Conditions. For given α, ω, d, z0 (t) and initial conditions the trajectories
are uniquely defined and we can denote
Put also
1
d∗ = d∗ (vmax , amax ) = (amax + αvmax )
ω2
Consider firstly the simplest initial conditions
dzk
zk (0) = −kd, (0) = v, k ≥ 0 (4)
dt
Stability and Admissible Densities in Transportation Flow Models 291
Theorem 1. Assume (4) and α > 2ω > 0. Then for any chosen “safe distance”
parameter d > d∗ in the Eq. (3) the following bounds hold
D∗ = max{A, d∗ },
where
αa + 2c α2
A= ,γ= − ω 2 , a = max{|a − d|, |b − d|}.
2γ 4
∗
Theorem 2. If α > 2ω, αa−2cα−2γ > 2 ,
a+b
2 < d < α−2γ then for any initial
a+b αa−2c
conditions (5) and any smooth function z0 (t), satisfying (2), there exists open
subset D ⊂ R such that for any choosen “safe distance” parameter d ∈ D in the
Eq. (3), the following bounds hold
I ≥ (d − D∗ ) > 0, S ≤ d + D∗ .
The proof of both theorems is based on the analysis of the chain of equations
for xk = rk − d, k > 0
xk (t) + αxk (t) + ω 2 xk (t) = ω 2 xk−1 (t), k = 1, 2, ...
where
1
x0 (t) = (z0 (t) + αz0 (t))
ω 2
Density and Currents. Let n(t, I) be the number of units on the interval I ⊂ R at
time t and n(T, x) be the number of units passing the point x in the time interval
(0, T ). Then the density and the current through some point x are defined as
follows (|I| is the length of I)
n(t, I)
μ(t) = lim
|I|→∞ |I|
n(T, x)
J(x) = lim
T
T →∞
if these limits exist. For fixed N define also mean length of the chain of cars
0, 1, ..., N
z0 (t) − zN (t)
LN (t) = .
N
292 A.A. Lykov et al.
Theorem 3. Under the conditions of Theorem 2 assume also that the initial
conditions are such that the following finite limits exist
and moreover
1 dL
L(t) = L(0) + (1 − e−αt ) (0)
α dt
Theorem 4. Let α = 0, then for the initial conditions (4) for k > 0 and for
z0 (t) = tv + sin ω t, v > 0, ω = 0 and k 2 we have
inf rk (t) = −∞
t0
1. for any α > 0 there exists ω > 0 and constants q+ > 1, μ+ > 0, c+ > 0,
so that for any d > 0
c+ k
rk (t) ∼ √ q+
k
as t = μ+ k, k → ∞;
2. for any α > 0 there exists ω > 0 and constants q− > 1, μ− > 0, c− < 0,
so that for any d > 0
c− k
rk (t) ∼ √ q−
k
as t = μ− k, k → ∞.
Stability and Admissible Densities in Transportation Flow Models 293
zk (t) = ck + vk t
It is easy to see that in our case at time t any density μ(t) is admissible iff
μ(t) ≤ d−1
and we again consider networks where the velocities and safe distance are the
same for all roads
Proposition 1.
1. Assume there are M < ∞ roads with only one common cross. In this case
the graph G consists of one vertex only. Then at time t any density μ(t) is
admissible iff
μ(t) ≤ (M d)−1 (6)
294 A.A. Lykov et al.
2. Assume there are M roads and let L be maximal multiplicity of their inter-
sections. Assume also that the graph G has no cycles. Then at time t any
density μ(t) is admissible iff
μ(t) ≤ (Ld)−1
zk,i (t) = zk,i (0) + vk t, zk,i−1 (0) − zk,i (0) = dk =⇒ zk,i−1 (t) − zk,i (t) = dk
Proposition 2. Consider two roads with one cross. Call the transportation sta-
ble if distances between any two units are greater or equal to some D0 > 0
uniformly in t. Then:
If JJ12 is not rational, then such transportation cannot be stable.
If JJ12 = nn12 for some integers n1 , n2 such that (n1 , n2 ) = 1. Then stable
transportation exists iff
n1 n2 1
+ <
d1 d2 D0
Remark 1 (Control types). To avoid instability of Proposition 2 some control is
necessary. There are two possibilities for the control. The first one we considered
in Sect. 2 - local internal control, that is depending only on distances between
cars. Second type is the control which forces velocities of cars to change in certain
points of the network. The mostly used such control is the organization of traffic
lights where the cars should stand still for some time. Other control types are
also known, see [1,4–8], and it could be interesting to find general classification
of control types.
References
1. Blank, M.: Ergodic properties of a simple deterministic traffic flow model. J. Stat.
Phys. 111, 903–930 (2003)
2. Prigogine, I., Herman, R.: Kinetic Theory of Vehicular Traffic. Elsevier, N.Y. (1971)
Stability and Admissible Densities in Transportation Flow Models 295
3. Helbing, D.: Traffic and related self-driven many particle systems. Rev. Mod. Phys.
73, 1067–1141 (2001)
4. Feintuch, A., Francis, B.: Infinite chains of kinematic points. Automatica 48,
901–908 (2012)
5. Hui, Q., Berg, J.M.: Semistability theory for spatially distributed systems. In: Pro-
ceedings of the IEEE Conference on Decision and Control, January 2009
6. Jovanovic, M.R., Bamieh, B.: On the Ill-Posedness of certain vehicular platoon
control problems. IEEE Trans. Autom. Control 50(9), 1307–1321 (2005)
7. Melzer, S.M., Kuo, B.C.: Optimal regulation of systems described by a countably
infinite number of objects. Automatica 7, 359–366 (1971). Pergamon Press
8. Swaroop, D., Hedrick, J.K.: String stability of interconnected systems. IEEE Trans.
Autom. Control 41(3), 349–357 (1996)
9. Malyshev, V.A., Musychka, S.A.: Dynamical phase transition in the simplest mole-
cular chain model. Theoret. Math. Phys. 179(1), 123–133 (2014)
Econometric Models of Controllable Multiple
Queuing Systems
Alexander Mandel(B)
1 Introduction
theory by looking into the possibility of stating and resolving control problems
for so-called conflict queuing systems.
The end of the 1980s saw an increasing interest in controllable queuing sys-
tems with the advent of an essentially new range of applied systems that could
be regarded and analyzed as queuing systems. These were computing and data
processing computer networks. Authorities in the queuing system theory con-
centrated their efforts on developing specific (in particular, network) applied
service systems. In Russia followers of Professors V.V. Rykov, V.M. Vishnevskii
and V.A. Kashtanov [5–7] made the largest contribution to this field of research.
As a consequence, over approximately two decades from 1985 to 2005
numerous models of controllable queuing systems that could provide an ade-
quate description, analysis and optimization of socio-economic systems remained
neglected. What is important is that purely econometric analysis of such systems
could provide a thorough and proper recognition of all economic indices that
are essential in evaluating the performance of such probabilistic and stationary
systems.
Papers [8, 9] pioneered in this research and this paper builds on their contri-
bution. We will take up two models. The one, referred to as stationary, assumes
that the input flow is Poisson with a variable corresponding to a discrete time
homogeneous Markov chain rate. In the other, referred to as nonstationary, the
input rate varies with the specified time function. Control of such QS reduces to
developing a spare service device switching strategy or a strategy of disconnect-
ing main service devices (turning them to spare).
decisions are made to switch on new service devices of to switch off some service
devices. The interval τ will be referred to as a “step”.
Let the input flow rate λ at fixed times tn = nτ, n = 1, 2, . . . , N , jumps from
λi to λj with probabilities pij (m1 ), i, j = 1, 2, . . . , L1 . It is essentially possible
that the probabilities also depend on n. Then pij (m1 ) should be replaced with
pij (m1 , n)). Consequently, at each interval of duration τ the demand input flow
rate is constant and described by a model of random quantity that takes on
values from a finite set Λ = λ1 , λ2 , . . . , λL assuming that λ1 < λ2 < · · · < λL .
We will further assume that M μ > λL , where μ is the service rate in one main
service device while M = m1 + m2 .
At times tn spare service devices can be switched into the set of main ones
and the inverse is true; in other words, spare service devices can be made main
ones, or devices can be switched on or off. This queuing model is ideologically
very much akin to the reliability model proposed in the 1960s and referred to as
dynamic reservation [10,11], even though, as will be shown below, it is a much
closer relation to two-level inventory control strategies of [12].
We will take up the case of setting stationary modes in a queuing system,
i.e. the case wherein in every interval of duration τ the number of demands
that can arrive and be served is sufficient for the queuing system to get into a
stationary mode in the probabilistic sense at each of its steps. For this purpose
it is sufficient that:
μτ >> 1. (1)
In this case it is also evident that λτ >> 1.
The assumption of the stationary mode imposes a fairly strong constraint on
the proposed model.
Firstly, next inequalities: ρi = λi /μM < 1, λi ∈ Λ, must hold. In effect, the
system can be stable only with a flow for which the condition holds:
1
The dependence of the probability of the jump-wise change in the input flow rate
on the number of the service devices m1 is linked to the need of accounting for the
effect of the servicing quality (in this case it depends on the time that the customer
has to wait for the service to start) on the customer demand, which is exactly what
is described by the input flow rate.
Econometric Models of Controllable Multiple Queuing Systems 299
Indeed, in statei (in line with the value of the input rate) the stationarity
(i)
condition ρi = λi /μm1m < 1, must hold, thence we have:
(i) λi (i) λi
m1m > ⇒, m1m = + 1 ∀i = 1, N . (4)
μ μ
Consequently, the number of devices operating in the state i may vary from
(i)
m1m to M.
where E is the mean value, kt, m1m , λi is the number of demands in the queue
to the QS with parameters m1m and λi at time t, while k(m1m , λi ) is the average
length of the queue to the QS with parameters m1m and λi . By virtue of the sta-
tionarity assumption, the integration is made over the interval [0, τ ] rather than
300 A. Mandel
over the current interval [tn , tn+1 ]. According to [1] we have for the stationary
case:
(m1m ρi )m1m ρi
k(m1m , λi ) = π0 , (7)
m1m !(1 − ρi )2
−1
m1m (m1m ρi )i (m1m ρi )m1m
where ρi = μmλi1m , while π0 = i=0 i! + m1m !(1−ρi ) is the station-
ary probability that there are no demands in the system.
The parameter d characterizes the indirect costs of the QS due to the longer
demands stay in the system which reduces the preference of the system by the
clients. A mathematical model may be devised that would relate the level of
preference loss to the average time of demand staying in the queue. That model
could lead to a realistic estimate of the variable d. This paper will not address
the problem. A remark is, however, in order that in most cases the value of the
variable d is 1,5-2 orders of magnitude lower than that of the above variable h.
The per step cost of operating main and spare service devices totals:
where λi ∈ Λ.
Just s steps before the end of the planning period we have
where the indices of step time boundaries use a structure of “inverse” time in
steps to the end of the planning period [0, T ]. In essence, the plot of input flow
rates is described as a totality of indices J = {jn }N
(n=1) .
Now the system’s state at every time tn is described as the number i, of
demands currently inside the QS. We will assume that the inequality (3) still
holds where this time λmax = λ(M ) . In the state jn (associated with the value
of the input flow rate), as noted above, the stationarity condition ρ(jm ) =
λ(jm ) /μM < 1, λ(jm ) ∈ Λ , must hold, whence:
λ(jm ) λ(jm )
∀λ(jm ) ∈ Λ .
(j ) (j )
m1mm > ⇒ m1mm = +1 (13)
μ μ
In effect, in the state jn the number of operating main devices may take on
(j )
values from m1mm to M . Because the case is non-stationary, in further discussion
we will discard the constraint (13).
Introduce the quantity C (1) (λ(j) , i, m1 , m1m ), which is equal to the averaged
income value at one step of duration τ if at the start of the step the input flow
rate value λ(j) sets in and there are i demands in the QS and the number of
main devices with which the QS arrives at this step is equal to m1 and a control
decision is made to switch on m1m main devices at the time when the step starts.
While doing so, formulae (5) and (8) remain in force while formulae (6), (9), (10)
and (11) need adjustment.
Indeed, no stationary mode arises at every step now and so the formulae for
computing the responses of the stationary mode do not hold. This applies to the
2
For that purpose the step length τ must be made small enough.
302 A. Mandel
formulae for the number of demands processed at the step and for the average
queue length. The key elements of the calculation become now the probabilities
of the QS state at the latest step, or the probabilities Pi (t) that at time t (recall
that t ∈ [tN −n+1 , tN −n ]) there are i demands in the QS.
A set of differential equations for the probabilities Pi (t) [1]:
dP0 (t)
for i = 0 : dt = −λP0 (t) + μP1 (t).
for 1 ≤ i < m1m : dPdti (t)
= λP(i−1) (t) − (λ + iμ)Pi (t) + (i + 1)μP(i+1) (t). (14)
dPi (t)
for i ≥ m1m : dt = λP(i−1) (t) − (λ + m1m μ)Pi (t) + m1m μP(i+1) (t),
with an initial condition
1 if j = i,
Pj (t) = (15)
0 if j = i.
To avoid resolving an infinite-dimensional set of differential equation (14)–
(15), we use an approximating set of finite dimension I:
(I)
dP0 (t) (I) (I)
for i = 0 : dt = −λP0 (t) + μP1 (t).
dPi (t)
for 1 ≤ i < m1m : dt = λPi−1 (t) − (λ + i)Pi (t) + (i + 1)Pi+1 (t).
(I)
dPi (t) (I) (I) (I)
for m1m ≤ i ≤ I: dt = λPi−1 (t) − (λ + m1m μ)Pi+1 (t) + m1m μPi+1 (t),
(16)
with an initial condition
(I) 1 if j = i,
Pj (t) = (17)
0 if j = i.
Assume that we have a “solver” of the equation set (16)–(17). Denote a
solution to the set (16) with an initial condition (17) as P(I) (i, τ ). The second
argument τ “reminds” one that the set (16)–(17) is resolved in the interval
[0, τ ]. Now, while assuming that we have an approximate solution to the set
(I)
(14)–(15) in the form of probabilities Pi (t), i = 1, 2, . . . , I, t ∈ [0, τ ], proceed
to computation of the other characteristics.
In further discussion the computation will be performed in inverse time s and
so the plot of the input flow rate will be represented by the values of λ(N −s) , s =
1, 2, . . . , N . Re-denote these rates as λ(N −s) = λmod . Let Πs∗ (λ, i, m1 ) be the
(s)
maximal value of the averaged income over an interval that starts s steps before
the end of the planning period [0, T ], s + n = N with the value λ (naturally
(s)
λ = λmod ) of the input flow rate, the number of demands inside the system i
and m1 of the switched on (before a control decision is made to switch on the
m1m main devices. Dynamic programming equations are derived below for the
functional Πs∗ (λ, i, m1 ) but this time in disregard of the condition (13).
Evidently that one step before the end of the planning period and with a
random value of the input flow rate λ the value of the functional Π1∗ (λ, i, m1 )
introduced above will look like
Proceeding now to adjustment of formula (6) for queuing cost. The average
queue length k(t, m1m , λ) at time t ∈ [ts , ts−1 ], where s is inverse time may be
represented in the form:
I
(I)
k(t, m1m , λ) = (i − m1m )Pi (t). (19)
i=m1m +1
The income from serving the demands over the interval [ts , ts−1 ] will be
estimated by using the Wald identity [2]. The average number of demands served
over the interval [ts , ts−1 ], will be estimated by an integral over the instantaneous
service rate at time t which is equal to
m1m −1 I
(I) (I)
μ(t) = μ Pi (t)i + m1m μ Pi (t) (21)
i=0 i=m1m
Consequently, the income from serving demands over the interval [ts , ts−1 ] is
ts−1
evaluated as h μ(t)dt . This expression makes it possible to rearrange formula
ts
(9) for the average net income within one step as
ts−1
C (1)
(λ, i, m1 , m1m ) = h μ(t)dt − Cswitch − Cqueue − Crunning , (22)
ts
where three cost components in the right-hand part of Eq. (22) are computed by
formulae (5), (20) and (8), respectively.
Now, s steps before the end of the planning period we have:
I (I) ∗ (s−1)
+ Pj (t)Πs−1 (λmod , j, m1m )} (23)
j=o
where ∀s ∈ 2, N − 1.
6 Conclusions
Models are proposed for control of controllable multiple queuing systems. Those
models are helpful in analyzing and optimizing the functioning of various socio-
economic systems such as booking air tickets and hotels, trading and purchasing,
public transport, billing and numerous others. The simple input flow rate varies
304 A. Mandel
References
1. Gnedenko, B.V., Kovalenko, I.N.: Introduction to Queuing Theory. Nauka Pub-
lisher, Moscow (1966). (in Russian)
2. Saaty, T.L.: Elements of Queuing Theory with Applications. McGraw-Hill,
New York (1961). no ISBN (translated into Russian, Spanish and German)
3. Rykov, V.V.: Controllable queuing systems. probability theory. mathematical sta-
tistic. theoretical cybernetic, vol. 12, pp. 43-153 (1975) (in Russian)
4. Neymark, Y.I.: Dynamic Systems and Controllable Processes. Nauka Publisher,
Moscow (1978). (in Russian)
5. Golysheva, N.M., Fedotkin, M.A.: Conflict flows cyclic control under critical length
queues birth-death process. Automation and Remote Control, No.4 (1990)
6. Vishnevskii, V.M.: Theoretical Foundations of Computer Networks Design. Tech-
nosphera Publisher, Moscow (2003). (in Russian)
7. Zaharov, P.P.: Computer-aided software complex development for quality and effi-
ciency research of technical and controllable queuing systems models. Engineering-
Science C and Dissertation. MGIEM, Moscow (2006) (in Russian)
8. Barladyan, I.I., Kuznetsov, A.V., Mandel, A.S.: Critical parameters analysis and
simulation of controllable queuing system. Problemy upravleniya, vol. 6, pp. 21–25
(2007) (in Russian)
9. Kuznetsov, A.V., Mandel, A.S., Tokmakova, A.B.: One model of controllable queu-
ing system. Problemy upravleniya, vol. 6, pp. 39–43 (2007) (in Russian)
10. Raikin, A.L.: Reliability Theory Topics for Technical Systems Design, p. 256. Sov.
Radio Publisher, Moscow (1967) (in Russian)
11. Mandel, A.S., Raikin, A.L.: Spare units switching optimal plan construction.
Automation and Remote Control, vol. 5 (1967)
12. Headly, J., Whiting, T.: Inventory Control Systems Analysis. Wiley, New York
(1967)
A Cluster Caching Rule in Next Generation
Networks
N. Markovich(B)
1 Introduction
The paper is devoted to probabilistic aspects of caching. A caching policy serves
to keep popular contents inside a short memory unit called ‘cache’ to be able
to access them as soon as possible. We consider it from new perspectives using
achievements of extreme value theory.
Let C be the size of a cache and d1 , ..., dM be a catalog of documents. Since
the popularity of contents may change over time and place, the hitting of docu-
ments inside the cache has to be changed, too. Cache filling depends on the input
(the arrival) process of document requests, the cache size and the replacement
policy. The latter may be controled by the time that is allowed for the docu-
ment to be in the cache. This called TTL (the Time-to-Live) may be individual
for each document depending on its popularity [1] or cache dependent [2], or
the TTL may be fixed, [3]. If a requested document is found in the cache then
the cache content remains unchanged. Then we say that the document request
hits the cache. If the document is new, i.e. it cannot be found in the cache,
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 305–313, 2016.
DOI: 10.1007/978-3-319-30843-2 32
306 N. Markovich
then the missing content is brought in from the outside world (a long memory).
In this case, we say that the document request misses the cache. The previous
cache content may remain unchanged in this case too if the cache was not full.
Otherwise, the new document evicts some content from the cache according to
the replacement rule.1
There are several identification problems in the caching such as size and
location of caches, statistical estimation of the content popularity, a convenient
modeling of the inter-request time (IRT) process, the optimality evaluation of
the replacement rule, e.g., by hit/miss probabilities as well as the occupancy
and the utilization of the cache. The most popular replacement rule is the Least-
Recently-Used (LRU) caching, [3,4], where the requested document hits the first
position of the cache. The object at the last position may leave the cache if the
requested document is new or it stays longer otherwise.
Despite of an easy idea, the analytical analysis of the LRU is difficult. Usually,
it is assumed for simplicity that the content size is deterministic despite it is
naturally random. However, one can combine several content units into chunks
of a constant size. Such chunks are assumed to be claimed.
The requesting point process of the object is often assumed to be a renewal
Poisson process and the IRTs are then independent exponentially distributed,
see [1,4] among others. A superposition of renewal Poisson processes is also
renewal Poisson. This simplifies the use of cache trees where superposed request
processes from leaf-caches arrive to the root-cache, [1]. Generally, the super-
position of two renewal processes is not a renewal process. Markov modulated
processes are closed under superposition. Hence, Markov and semi-Markov mod-
ulated processes are typically used as alternative models of arrival request
processes, [1,5,6]. In contrast to renewal Poisson processes such superposed
request processes are correlated.
Usually, the probability of the content popularity is modeled by Zipf’s law
that has a Pareto tail, [7,8]. In [9] the mixture of Zipf and heavy-tailed Weibull
distributions is found as an appropriate model of the web content popularity. In
[4] real web traces from the National Laboratory for Applied Network Research
(NLANR) were analyzed. It was found that about 70 % of all documents are
one-time requested. The caching of such documents is not reasonable. However,
the LRU and TTL rules allow to cache such documents.
Assuming that the document popularity can be modeled by a generalized
Zipf’s law, in [5] it was derived that the cache missing probability is asymptot-
ically, for sufficiently large cache sizes, the same for the LRU replacement both
with dependent and independent identically distributed IRTs. In [6] similar result
was proved for the Least-Frequently-Used (LFU) rule which only caches the most
popular objects. The IRTs of each object may be dependent and heavy-tailed
distributed which reflects the heavy-tailed popularity of the documents.
According to our new caching rule called Cluster Caching Rule (CCR), only
clusters of frequently requested contents whose popularity exceeds a sufficiently
1
If the document size is fixed then a new document evicts one document from the
full cache.
A Cluster Caching Rule in Next Generation Networks 307
large threshold may hit the cache. This is similar to TTL and LFU rules, [10]
but there are novelties based on the extreme value theory. The latter allows
us to explain the impact of the dependence and heavy-tailed quantities on the
caching. Due to dependence and a possibly heavy-tailed distribution of IRTs of
the object, the popularity and IRT processes build clusters of exceedances, see
Fig. 1. Such assumptions like (1) the content size is constant; (2) the IRT process
is renewal Poisson; (3) IRTs are independent are avoided.
We consider clusters of exceedances of a popularity process of random sized
content whose IRTs may be statistically correlated. Inter-arrival times of docu-
ments can be heavy-tailed. The clusters provide objects to hit/miss the cache.
We obtain the cache utilization and occupancy by means of clusters and explain
the impact of correlated requests on the fault probability from perspectives of
clustering. The CCR allows to recommend the cache size.
We focus on clusters of exceedances of the popularity process. The popularity
may change over time. When the next cluster of the most popular documents is
requested, it is sent to the cache. The former cache content is evicted. If some
popular content is requested repeatedly then it remains in the cache without
replacement. In case, the cache size is smaller than the cluster size, the threshold
u may be increased until the cluster size becomes smaller or equal to the cache
size. The cache content may be changed completely by a new cluster. In our
approach, the document size can be variable, i.e. a random variable (r.v.). We
concern only probabilistic aspects of fitting the cache avoiding statistical and
combinatorial methods.
308 N. Markovich
where M1,j = max{X2 , ..., Xj }, M1,1 = −∞, L1,j = min{X2 , ..., Xj }, L1,1 =
+∞, Fig. 1.
In [11] asymptotically equivalent distributions of T1 (un ) and T2 (un ) are
derived for sequence of increasing thresholds {un }. It was proposed to take suf-
ficiently high quantiles xρn of the level (1 − ρn ) (ρn → 0 as n → ∞2 ) of the
process Xt as un . If specific mixing conditions are fulfilled uniformly in j ∈ [a, n],
a = a(n) → ∞, n → ∞ then geometric-like models.
(j−1)θ
P {T1 (xρn ) = j} ∼ ρn (1 − ρn )(j−1)θ , P {T2 (xρn ) = j} ∼ qn (1 − qn )
(2)
2
This follows from (1) since ρn = 1 − F (xρn ) ∼ τ /n.
A Cluster Caching Rule in Next Generation Networks 309
It is remarkable that models (2) are valid for any distribution of the underlying
process {Xt } if the dependence structure (3) holds.
3 Main Results
Let {Xi }1≤i≤M be a stationary process of the content popularity with marginal
cumulative distribution function F (x) and Mn = max{X1 , ..., Xn }. We assume
P {X1 = xF } = 0, where xF = sup{x : F (x) < 1} is the right end-point of F (x).
This indicates that xF is infinite and F (x) is heavy-tailed. Let {Yi } be iid sizes
of documents from the catalog with finite mean EYi = α1 .
Let {τn }n≥1 and {τi,n }n≥1 be stationary processes of inter-arrival times
(IATs) of documents and IRTs of the ith document, i = 1, .., M , respectively.
The frequency of requests changes over time. We j determine the popularity of
the ith document at its jth request time Ti,j = n=1 τi,n by
Xi = j/NTi,j , i = 1, ..., M, j ≥ 1,
where NTi,j is the total number of requests in time interval [0, Ti,j ].
3
fn ∼ gn implies limn→∞ fn /gn = 1.
310 N. Markovich
Let us consider the cache utilization for fixed and random content sizes. Let
Yi be fixed. The utilization may be determined by the ratio T2∗ (xρn )/C of the
cluster and the cache sizes. Clusters may contain changing popularity of the
same documents. Since ET2 (u) ≈ 1/θ holds we get
T2 (u)
T1 (u)
ST2 (u) = Yi , ST1 (u) = τi . (7)
i=1 i=1
E(ST2∗ (u) /C) ≤ E(ST2 (u) /C) = E(T2 (xρn ))E(Yi )/C ≈ α1 /(θC) (8)
A Cluster Caching Rule in Next Generation Networks 311
Thus, the mean utilization may reach 100 % if C = α1 /θ. The cache size
requires a statistical estimation of the extremal index θ. Since θ is a function of
the threshold u, one can use (4) to determine C. Equation (6) can be rewritten as
The cache occupancy shows how long an object stays in cache within some
time interval. Let C(t) denote a binary process that shows the presence of the
ith document in the cache. Denoting independent renewal processes of IRTs of
individual objects and their TTLs as {τi,k }k≥1 and {Ti,k }k≥1 , respectively, in
[1] the following formula of the cache occupancy
tn
n
n
lim C(s)ds/tn = lim min{τi,k , Ti,k }/ τi,k (9)
n→∞ 0 n→∞
k=1 k=1
n
was proposed. Here, tn = k=1 τi,k is the time of the nth request of the object
and C(s) is a binary process indicating the presence of the object in cache.
Equation (9) corresponds to the TTL rule with R-policy when the stopping time
to be in cache is equal to min{t : τt > Tt }.
The CCR-approach is different from (9). Once the object hits the cache, it
will stay there at least the time equal to the inter-cluster duration ST1 (u) . Thus,
assuming that the IATs {τi } of the documents are iid r.v.s with β1 = E(τi ) < ∞
and the mutual independence of τi and T1 (u), we obtain from Wald’s equation
and (5), (7)
with ρn ∼ τ /n. More precisely, the cache occupancy of the ith object may be
longer or equal to ⎛ ⎞
T1,j
Li,T
(u)
T2,j (u)
Oi ≥ ⎝ τk + τk ⎠ ,
j=1 k=1 k=k∗
where k ∗ = min{t : Xt ≥ u} is the first hitting time of the cache by the ith object
located in the cluster, Li,T is the number of consecutive clusters containing the
ith document and, T1,j (u) and T2,j (u) are inter- and cluster sizes, respectively,
at time interval [0, T ].
4
This follows from (1) since 1 − F (xρn ) = ρn = 1 − qn ∼ τ /n as n → ∞.
312 N. Markovich
Acknowledgments. The author was partly supported by the Russian Foundation for
Basic Research, grant 13-08-00744 A.
References
1. Berger, D.S., Gland, P., Singla, S., Ciucu, F.: Exact analysis of TTL cache
networks: the case of caching policies driven by stopping times. In: The ACM
International Conference on Measurement and Modeling of Computer Systems,
SIGMETRICS 2014, pp. 595–596 (2014)
A Cluster Caching Rule in Next Generation Networks 313
2. Foback, N.C., Nain, P., Neglia, G., Towsley, D.: Analysis of ttl-based cache net-
works. In: IEEE VALUETOOLS, pp. 1–10 (2012)
3. Friecker, C., Robert, P., Roberts, J.: A versatile and accurate approximation for
LRU cache performance. In: Proceedings of ITC, pp. 1–8 (2012)
4. Che, H., Tung, Y., Wang, Z.: Hierarchical web caching systems: modeling, design
and experimental results. IEEE JSAC 20(7), 1305–1314 (2002)
5. Jelenković, P.R., Radovanović, A.: Least-recently-used caching with dependent
requests. Theoret. Comput. Sci. 326(1–3), 293–327 (2004)
6. Jelenković, P.R., Radovanović, A.: Asymptotic optimality of the static frequency
caching in the presence of correlated requests. Oper. Res. Lett. 37(5), 307–311
(2009)
7. Clauset, A., Shalizi, C.R., Newman, M.E.J.: Power-law distributions in empirical
data. SIAM Rev. 51(4), 661–703 (2009)
8. Newman, M.E.J.: Power laws, Pareto distributions and Zipfs law [cond-mat.stat-
mech] (2006). arxiv:cond-mat/0412004v3
9. Imbrenda, C., Muscariello, L., Rossi, D.: Analyzing cacheable traffic in ISP access
networks for micro CDN applications via content-centric networking. In: ACM
SIGCOMM Information Centric Networks (ICN), September 2014
10. Lee, D., Choi, J., Kim, J.-H., Noh, S.H., Min, S.L., Cho, Y., Kim, C.S.: LRFU:
a spectrum of policies that subsumes the least recently used and least frequently
used policies. IEEE Trans. Comput. 50(12), 1352–1362 (2001)
11. Markovich, N.M.: Modeling clusters of extreme values. Extremes 17(1), 97–125
(2014)
12. Leadbetter, M.R., Lingren, G., Rootzén, H.: Extremes and Related Properties of
Random Sequence and Processes. Springer, New York (1983)
13. Beirlant, J., Goegebeur, Y., Teugels, J., Segers, J.: Statistics of Extremes: Theory
and Applications. Wiley, Chichester, West Sussex (2004)
14. Markovich, N.M.: Clusters of extremes: models, estimation and applications (2015,
to submitted)
On Initial Width of Contention Window
Influence on Wireless Network Station
IEEE 802.11 Characteristics
of the contention window may not exceed the maximum value established by the
standard. For all physical layers and methods of modulation, the IEEE 802.11
standard has established the maximum width of the contention window equal to
Smax = 1024 [2] The number of a selected slot shall be assigned to the backoff
interval counter to , after which the countdown of slot intervals begins. At the
end of each slot interval, the backoff interval counter shall decrement as long as
medium is idle. If the medium is determined to be busy at any time during a
backoff slot, then the backoff procedure is suspended. Decrementing is resumed
when the medium is idle again. Transmission shall commence when the backoff
interval counter reaches zero (to = 0). When the transmission is completed, the
sender waits for a acknowledgement during the time tout , after which it is con-
sidered that a conflict has occurred, and stations having got into such conflict
increase the n value by one, and the actions targeted at data transmission are
repeated. The width of the contention window is doubled with each attempt of
data frame transmission, until the maximum value is achieved; and the width
of the contention window remains equal to Smax with each subsequent attempt
of data frame transmission. After successful transmission, the window width
obtains the initial value S0 .
Thus, the wireless access technology, due to lack of possibility to detect col-
lisions in a communication medium, has three significant differences from the
random access method implemented in the wired medium. Firstly, the wire-
less transmission method employs the mechanism of positive feedback (positive
acknowledgements). Secondly, in contrast to the random access method, in wired
networks the WiFi technology employs the random delay mechanism as early as
during the first transmission. And at last, the wireless access protocol employs
the mechanism of “suspension” of the delaying timer from the time of detection
of the medium occupation until expiration of the random delay timer.
Let us analyze the operation of a wireless local area network until the first
error-free data frame transmission with obtained acknowledgement on successful
delivery of data. Let us suppose that the wireless LAN contains K stations which
are data sources. Consider that all the sources are independent and equal, and
always have data frames for sending, and all interval spaces are expressed in slot
intervals tc . Let all the stations exchange frames of equal sizes. Then, according
to the sequence of protocol actions, the elementary cycle of data frame transfer
to the recipient will be determined by the size of the interframe space tm , random
delay period to , duration of “suspension” of the random delay timer tz , time of
data frame transmission tk , and the value of time-out for expecting a positive
acknowledgement tout , which consists of a short interframe space plus the time
of transmission of a positive acknowledgement [2,4]. The average time of data
frame transmission T (K, N0 ) consists of the weighted sum of average periods of
316 P. Mikheev and S. Suschenko
waiting for failed transmissions and the time of successful transmission [6]:
∞
N −1
T (K, N0 ) = d + Nd + t(n, K, N0 ) + τ (N, K, N0 ) f (N, K, N0 ). (1)
N =0 n=0
N −1
f (N, K, N0 ) = P (N, K, N0 ) π(n, K, N0 ).
n=0
Along with the average time of data frame transmission, one of the main
indicators showing the efficiency of functioning the data transfer network is the
throughput performance. In the case under analysis, we will look for an individ-
ual throughput performance, the standardized value of which shall be determined
as a ratio between the time necessary for data frame transmission tk and the
average time of data frame transmission T (K, N0 ):
tk
C(K, N0 ) = . (2)
T (K, N0 )
π(n, 2, N0 )
⎧S −1
⎪
⎪
0 i
⎨ i=0 p0 (i) j=0 f0 (j)Li−j ,
⎪ n = 0;
= Sk −1 Sk −1
⎪
⎪
n i Sn −1
⎪
⎩ Ek (n) pn (i) fk (j)Li−j + pn (i) fk (j)Li−j , n ≥ 1.
k=1 i=0 j=0 i=Sk j=0
(3)
On Initial Width of Contention Window Influence on Wireless Network 317
The coefficients Ek (n) in the relation (3) are the probabilities that on the n-th
repeated transmission by the A station, the B station will be in the condition of
the k-th repeated transmission:
⎡
Ek (n − 1)
n−1 k −1
S
i−1
E1 (1) = 1; E1 (n) = ⎣ pn−1 (i) fk (j)Li−j
π(n − 1, 2, N0 ) i=1 j=0
k=1
⎤
Sn−1 −1 k −1
S
+ pn−1 (i) fk (j)Li−j ⎦ , n ≥ 2;
i=Sk j=0
Sk−1 −1
Ek−1 (n − 1) pn−1 (i)fk−1 (i)
i=0
Ek (n) = , n ≥ 2, k = 2, n.
π(n − 1, 2, N0 )
The average conditional times until failed and successful n-th attempt of
data transmission t(N, K, N0 ) and τ (N, K, N0 ) consist of the average duration
of random delay Ns (n) (average number of slots until the start of transmission)
and the average number of suspensions caused by medium capture by the B
station, Zt (n, N0 ) in case of failure and Zτ (n, N0 ) in case of success, respectively:
t(n, 2, N0 ) = Ns (n) + Zt (n, N0 )d, τ (n, 2, N0 ) = Ns (n) + Zτ (n, N0 )d.
318 P. Mikheev and S. Suschenko
S −1
Here Ns (n) = i=0 n
ipn (i) = (Sn − 1)/2, and the average numbers of suspen-
sions Zt (n, N0 ) and Zτ (n, N0 ) look similar:
Zt (n, N0 )
⎧S −1
⎪ 0
i−1
⎪
⎨ i=1 p0 (i) j=0 f0 (j)Mi−j ,
⎪ n = 0;
=
⎪
⎪ n S
k −1
i−1
Sn −1 S
k −1
⎪
⎩ Ek (n) pn (i) fk (j)Mi−j + pn (i) fk (j)Mi−j , n ≥ 1;
k=1 i=1 j=0 i=Sk j=0
(5)
Zτ (n, N0 )
⎧S −1
⎪
0
i−1
⎪
⎨ i=1 p0 (i) j=0 f0 (j)Vi−j ,
⎪ n = 0;
=
⎪
⎪ n S
k −1
i−1
Sn −1 S
k −1
⎪
⎩ Ek (n) pn (i) fk (j)Vi−j + pn (i) fk (j)Vi−j , n ≥ 1.
k=1 i=1 j=0 i=Sk j=0
(6)
The elements Mk and Vk are indicators of the average number of suspensions
of the delaying timer for the A station after selection of random delay with
the duration i on the n-th repeated transmission upon selection of the j-th slot
preceding to the i-th one by the competing B station (k is a difference between
j-th and i-th slots):
⎧ k ∞
⎪
⎪
⎨ f0 (m) (i + 1 + Mk−m )f0i (0), k = 1, S0 − 1, M0 = 0;
Mk = m=1 S0 −1
i=0
∞
⎪
⎪
⎩ f0 (m) (i + 1 + Mk−m )f0i (0), k = S0 , Sn − 1;
m=1 i=0
Vk
⎧∞
⎪ S
0 −1
k−1
∞
⎪ i
⎨ (i + 1)f0 (0) f0 (m) + f0 (m) (i + 1 + Vk−m )f0i (0), k = 1, S0 − 1;
i=0 m=k+1 m=1 i=0
=
⎪S
⎪ 0 −1
∞
⎩ f0 (m) (i + 1 + Vk−m )f0i (0), k = S0 , Sn − 1.
m=1 i=0
After inserting here the probabilities of fallout of delay duration f0 (m), we obtain
the following relations:
⎧ k
⎪
⎪ S0 S0
⎪
⎨ − 1 , k = 1, S0 − 1;
S0 − 1 S0 − 1
Mk =
⎪
⎪ S0
S0 −1
Mk−m
⎪
⎩ + m=1 , k = S0 , Sn − 1.
S0 − 1 S0 − 1
⎧ k
⎪
⎪ S0 − 2 S0
⎨ , k = 1, S0 − 1;
Vk = S0 − 1 S0 − 1
S0 −1
⎪
⎩ S0 + m=1 Vk−m , k = S0 , Sn − 1.
⎪
S0 − 1 S0 − 1
On Initial Width of Contention Window Influence on Wireless Network 319
(G(N0 ) + 1)tk
Cg (2, N0 ) = ,
T (2, N0 )
where G(N0 ) will be determined by the weighted amount of the average number
of suspensions of the delaying timer of the A station in expectation of failed and
successful transmissions, which are determined by the relations (5) and (6):
∞
N −1
G(N0 ) = Zt (n, N0 ) + Zτ (N, N0 ) f (N, 2, N0 ).
N =0 n=0
4 Numerical Results
The numeric research into the average time of data frame transmission by the
A station shows that the function (1) has a strongly manifested minimum at
the coordinate N0 (see Fig. 1) determining the initial size of the competition
window and, subsequently, the degree of scattering of stations by durations of
delays before the start of the competition procedure. For two competing sta-
tions, the minimum is reached at N0 = 4. It is obvious that the value N0 min-
imizing the average time of data frame transmission maximizes the individual
throughput (see Fig. 1). Moreover, as early as at the stage of formalization of
the task, the probability of capture of the communication medium by one of the
subscribers mentioned in [9,10] has become obvious. This effect manifests itself
especially strongly with small values N0 . The effect of capturing the communica-
tion medium causes discrimination-related individual indicators against a good
level of the general throughput performance of the network (see Fig. 1).
As early as at the first attempt of competition between two stations, capture
of the communication medium becomes possible (e.g. by the B station), and its
probability will be determined by the probabilities that for one of the stations
(B) the delay duration will turn out to be shorter than the duration of delay of
the other station (A); then the “succeeded” station (B) will have fallout of zero
duration, which will alternate with shorter delays than the residual value of the
station’s A delaying timer:
0 −1
S ∞
S0 −1
1 S0
Pz (0, 2, N0 ) = p0 (i)Li−1 f0k (0) = .
i=1
S02 S0 − 1
k=1
From this point, it is not difficult to see that the probability of capture is consid-
erably determined by the initial width of the contention window S0 (see Fig. 2).
After several conflicts, the possibility of capture for the “succeeded” station
becomes yet more probable.
320 P. Mikheev and S. Suschenko
300 0,4
200
0,2
100
0
0 2 4 6 8 10 N0
Fig. 1. Average time of data frame transmission, and individual and general through-
put performances
Pz (0, 2, N0 )
0,5
0,4
0,3
0,2
0,1
0 2 4 6 8 10 N0
The main reason for the effect of capturing the communication medium is the
protocol action — “suspension of delay”, because this results in a fact that after
a non-conflict transmission the station may capture the communication medium
for an infinitely long time, getting into the delay interval from 0 to the residual
value of the delay of other stations.
Another reason for an increase in the probability of capturing the commu-
nication medium by one of the subscribers after several conflicts, consists in
various sizes of the contention window for stations withdrawn from the conflict
and stations continuing resolution of the conflict in the condition of waiting for
expiration of delay time and suspension periods. After a positive resolution of the
conflict by one of the stations (or by several stations), the size of its contention
window is reduced in multiples down to the initial value S0 < Sn , which gives
this station a priority right in subsequent competition for the medium with
On Initial Width of Contention Window Influence on Wireless Network 321
5 Conclusion
The performed analysis is targeted at studying the method of carrier sense multi-
ple access with collision avoidance. Analytic correlations have been obtained for
probability timing characteristics of the competition process between two sta-
tions. The “medium capture effect” and the extreme dependence of operational
parameters on the initial contention window size have been revealed.
It has been suggested to change the parameters of the protocol procedure of
competition, ensuring prevention of the capture effect by saving high values of
individual and integral indices of operational speed.
It has been shown that the optimal initial width of the contention win-
dow (S0 ) is determined by the active size of the network (the number of compet-
ing stations), and it ensures almost uniform distribution of a jointly used time
resource of the medium between competing subscribers.
References
1. Tanenbaum, A.S., Wetherall, D.J.: Computer Networks, 5th edn, p. 960. Prentice
Hall, Boston (2010)
2. IEEE Std 802.11 — 2007, Revision of IEEE Std 802.11 — 1999. IEEE Std
802.11 — 2007, IEEE Standard for Information Technology, Telecommunications,
information exchange between systems, Local, metropolitan area networks, Spe-
cific requirements. Part 11: Wireless LAN Medium Access Control (MAC) and
Physical Layer (PHY) Specifications, p. 1184. IEEE Computer Society (2007)
3. Geier, J.: Wireless Networks First-step. Cisco Press, Indianapolis (2005). p. 192
4. Roshan P., Leary J.: 802.11 Wireless LAN Fundamentals, p. 312. Cisco Press,
Indianapolis (2004)
5. Bianchi, G.: Performance analysis of the IEEE 802.11 distributed coordination
function. IEEE J. Sel. Areas Commun. 18(3), 535–547 (2000)
6. Kustov, N.T., Suschenko, S.P.: Capacity of the random multiple access method.
Autom. Remote Control. 62(1), 76–85 (2001)
7. Hastings, N.A.J., Peacock, J.B.: Statistical Distributions in Scientific Work Series:
A Handbook for Students and Practitioners, p. 130. Butterworth, London (1975)
8. Prudnikov, A.P., Brychkov, Y.A., Marichev, O.I.: Integrals and Series: Special
functions, p. 751. Overseas Publishers Association, Amsterdam (1986)
322 P. Mikheev and S. Suschenko
9. Bononi, L., Conti, M., Donatiello, L.: Design and performance evaluation of dis-
tributed contention control (DCC) mechanism for IEEE 802.11 wireless local area
network. J. Parallel Distrib. Comput. 60(4), 407–430 (2000)
10. Vishnevsky, V.M., Lyakhov, A.I.: IEEE 802.11 wireless LAN: saturation through-
put analysis with seizing effect consideration. Cluster Comput. 5, 133–144 (2002)
Tandem of Infinite-Server Queues
with Markovian Arrival Process
1 Introduction
Tandem queues are used for modelling of modern telecommunications and data
processing systems [1] in such cases when it is need to perform a processing of
calls or data packages in the several stages. Most researchers consider such mod-
els for case of Poisson arrivals [2] or for some specific configuration of the tandem
[3–5], usually only with two stages. However, there are results [6] which prove
that the Poisson model can be adequate only in a few cases of modern telecom-
munication streams. Therefore, many researchers use more complex models of
the streams such as Markovian arrival process [7] or semi-Markov process [8].
We consider the tandem of queues with infinite number of servers at each
stage and total number of stages K ≥ 2. Infinite-server models are popular in the
queueing theory because they allow to obtain some results which can be applied
to real finite-server systems after some assumptions. More information about
infinite-server models can be found in [9]. In [10], we performed some studies
about how the results for infinite-server model can be applied to the models
with limited number of servers.
A model and a problem under study are described in the Sect. 2. In the
Sect. 3, we present a method for constructing Kolmogorov equations (they a
derived in the Sect. 4) which allow to solve the problem. In the Sect. 5, we make
the asymptotic solution of derived equations under condition of growth of arrivals
rate. Also in that section, we obtain a Gaussian approximation for the multi-
dimensional probability distribution of customers number at the tandem stages.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 323–333, 2016.
DOI: 10.1007/978-3-319-30843-2 34
324 A. Moiseev and A. Nazarov
Numerical results are presented in the Sect. 6 and they allow to determine the
boundaries of the approximation applicability.
Similar results was obtained for the case of renewal arrival process [11].
2 Mathematical Model
The object of the study is a tandem of K infinite-server queues which we call
‘stages’. A customer that comes into the system enters the first stage for service
with the cumulative distribution function B1 (x). After the service is complete
at the first stage the customer moves to the next stage for the service, and so on
while the service at the last K-th stage will be complete. After that the customer
leaves the system. Denote by Bk (t) for k = 1, . . . , K cumulative distribution
functions of service time at the respective stages of tandem.
Process of customers’ arriving is Markovian arrival process (MAP) [12,13]
with representation (D 0 , D 1 ) of order M . The matrix D = D 0 + D 1 is a gener-
ator of the underlying Markov chain of MAP. Denote the stationary distribution
of the chain states by row vector θ. This vector satisfies the following equations:
θD = 0,
(1)
θe = 1.
Here 0 is a row vector with zeros and e is column vector with entries all equal
to 1. The value
λ = θD 1 e (2)
is called as fundamental rate of the Markovian arrival process [7]. It characterizes
the average rate of the arrivals.
Denote by ik (t) the number of customers at the k-th stage of the tandem at
the instant t. The goal of study is to find the multi-dimensional joint probability
distribution of the tandem state i(t) = {i1 (t), . . . , iK (t)}.
T
0
.........................................................................
Fig. 1. Generating events in the screened streams on base of the customers arrivals
We will say that a customer, which comes to the system at the moment t <
T , generates an event in the k-th component of the screened process with the
probability Sk (t). And it doesn’t generate any event in the screened process with
probability S0 (t).
Denote by n(t) = {n1 (t), . . . , nK (t)} a counting process of the described
screened point process. Here nk (t) is a number of events generated in the k-th
component of the screened process before the instant t. The basic idea of the
multi-dimensional dynamic screening method is that the probability distribution
of customers numbers in the system i(t) and the probability distribution of value
of multi-dimensional counting process n(t) are coincide at the instant t = T :
4 Kolmogorov Equations
Denote by m(t) the state of the underlying Markov chain of the Markovian arrival
process. Let’s consider the (K + 1)-dimensional Markovian process {n(t), m(t)}.
Using the notation p(n, m, t) = P {n(t) = n, m(t) = m} and applying the for-
mula of total probability, after some derivations we can write the following sys-
tem of Kolmogorov differential equations for the probability distribution of the
process {n(t), m(t)}:
∂p(n, m, t)
M
= p(n, η, t)(D 0 )ηm
∂t η=1
M
K
+ p(n, η, t)(D 1 )ηm S0 (t) + p(n − ek , η, t)(D 1 )ηm Sk (t)
η=1 k=1
∂p(n, t) K
= p(n, t)D 0 + p(n, t)D 1 S0 (t) + p(n − ek , t)D 1 Sk (t) (6)
∂t
k=1
Using the system (6), we can write the following matrix differential equation for
the function h(u, t):
∂h(u, t)
K
= h(u, t) D 0 + D 1 S0 (t) + D 1 e Sk (t) .
juk
∂t
k=1
Taking into account formula (3), we reduce this equation to the form
∂h(u, t) K
ju
= h(u, t) D + D 1 e k − 1 Sk (t) (7)
∂t
k=1
h(u, t0 ) = θ. (8)
Tandem of Infinite-Server Queues with Markovian Arrival Process 327
5 Asymptotic Analysis
We can not directly solve the problem (7), (8) using a matrix exponent because
the matrices D and D1 are not commutative: D · D1 = D1 · D. So, here we
obtain an asymptotic solution of this problem under a condition of growth of
the fundamental rate of arrivals using the technique [15]. To do this, we make
in the Eq. (7) the substitutions N D 0 instead the matrix D 0 and N D 1 instead
the matrix D 1 . Here N > 0 is some parameter. The fundamental rate of such
arrival process is equal to N λ. If value of the parameter N grows infinitely, then
value of the fundamental rate of the arrival process grows infinitely too. Such
arrival processes we call as high-rate or high-intensive processes [11,14].
As a result of described transforms we obtain the following matrix differential
equation
1 ∂h(u, t)
K
ju
= h(u, t) D + D 1 e k − 1 Sk (t) . (9)
N ∂t
k=1
The initial condition for this asymptotic problem is the same as expression (8).
f 1 (w, t0 , ε) = θ. (12)
Denote an asymptotic solution of this problem under a condition ε → 0 by
f 1 (w, t):
f 1 (w, t) = lim f 1 (w, t, ε).
ε→0
f 1 (w, t)D = 0.
Comparing this equation with the first one in the system (1), we can draw a
conclusion that the function f 1 (w, t) can be represented in the form
Φ1 (w, t0 ) = 1. (15)
1. Sum all rows of the matrix Eq. (11) by multiplication on vector e and take
into account that
De = 0. (16)
2. Divide the result by ε.
3. Use the expression (14).
4. Perform an asymptotic transition ε → 0.
∂Φ1 (w, t) K
θ e = θΦ1 (w, t)D 1 e jwk Sk (t).
∂t
k=1
Taking into account that θe = 1 and using the expression (2), we reduce this
equation to the form
∂Φ1 (w, t) K
= Φ1 (w, t)λ jwk Sk (t).
∂t
k=1
Substituting this expression into the formula (14), we obtain the expression
for the function f 1 (w, t) in the form (13). The theorem is proved.
Tandem of Infinite-Server Queues with Markovian Arrival Process 329
Substituting this expression into the formulas (9), (8), we obtain the following
Cauchy problem for the function h2 (u, t).
⎧
⎪ K
⎪
⎪ 1 ∂h2 (u,t)
+ h2 (u, t)λ juk Sk (t)
⎪
⎨ N ∂t
k=1
K
(18)
⎪
⎪ = h (u, t) D + D ejuk
− 1 S (t) ,
⎪
⎪
2 1 k
⎩ k=1
h2 (u, t0 ) = θ.
Theorem 2. The asymptotic solution f 2 (w, t) of the problem (20) under a con-
dition ε → 0 is the following
⎧ ⎫
⎨ K t
K K t ⎬
(jwk )2 j 2 wk wν
f 2 (w, t) = θ exp λ Sk (τ ) dτ + κ Sk (τ )Sν (τ ) dτ
⎩ 2 2 ⎭
k=1 t0 ν=1
k=1 t0
(22)
where
κ = 2g(D 1 − λI)e. (23)
Here the row vector g satisfies the following matrix linear equation
gD = θ(λI − D 1 ). (24)
330 A. Moiseev and A. Nazarov
where g is some row vector and O(ε ) is a row vector that consists of the infin-
2
itesimals of the order ε2 .Let’s substitute the expansion (27) and the expansion
ejεwk = 1 + jεwk + O ε2 into the first equation of the problem (20). We obtain
the equality
K
K
K
θλ jεwk Sk (t) = θD + θD 1 jεwk Sk (t) + gD jεwk Sk (t) + O(ε2 ).
k=1 k=1 k=1
Performing here the asymptotic transition ε → 0 and taking into account the
Eq. (1), we obtain the following matrix equation for the row vector g:
gD = θ(λI − D 1 )
which coincides with expression (24).
Stage 3. Multiplying both parts of the first equation of the problem (20) by vector
2
e, using the expansion ejεwk = 1+jεwk + (jεw2 k ) +O(ε3 ), taking into account the
properties (16), (1) and the expressions (2), (23), under an asymptotic condition
ε → 0 we obtain the following linear differential homogeneous (on variable t)
equation for the function Φ2 (w, t)
K
∂Φ2 (w, t) (jwk )2 K K
j 2 wk wν
= Φ2 (w, t) λ Sk (t) + κ Sk (t)Sν (t)
∂t 2 ν=1
2
k=1 k=1
where the value of κ is defined by expression (23). Under the initial condition
(26), the solution of this equation is the following
⎧ ⎫
⎨ K t
K K t ⎬
(jwk )2 j 2 wk wν
Φ2 (w, t) = exp λ Sk (τ ) dτ + κ Sk (τ )Sν (τ ) dτ .
⎩ 2 2 ⎭
k=1 t0 ν=1
k=1 t0
Tandem of Infinite-Server Queues with Markovian Arrival Process 331
Substituting this expression into (25), we obtain the final form of the function
f 2 (w, t) as the expression (22). The theorem is proved.
Let’s perform a transition to the random vector i(T ) which is the main
object of the study. Denote by h(u) the characteristic function of the probability
distribution of the random vector i(T ). Due to the main formula of the multi-
dimensional dynamic screening method (4), we have the following approximation
h(2) (u) for the characteristic function h(u) under a condition that N is large
enough:
⎧
⎨ K T
(juk )2
h(u) ≈ h (u) = exp N λ
(2)
juk + Sk (τ )dτ
⎩ 2
k=1 t0
⎫
juk juν
K K T
⎬
+ Nκ Sk (τ )Sν (τ )dτ .
2 ⎭
k=1 ν=1 t0
6 Numerical Results
To determine the accuracy of the approximation (28) we have performed simula-
tions of the system evolution for numerous concrete examples and then compare
resulting empiric frequencies with probabilities of a distribution constructed on
base of the approximation (28). The accuracy of the approximation we charac-
terize by the Kolmogorov distance [16] which has the following form for discrete
distributions: x
d = max [p̃(l) − p(l)].
x≥0
l=0
Here p(l) (l ≥ 0) is the distribution of Gaussian approximation in discrete points
0, 1, 2, . . . , and p̃(l) is empiric distribution of customers number at the system
stage constructed on base of simulation results. Parameters of the model are
chosen in a such way to make the average number of customers at the considered
stage of the system be equal to N . So, we can see the common trend of the
approximation applicability bounds in relation to value of parameter N as the
average number of customers at the stage.
Obtained values of the Kolmogorov distance d for various values of the para-
meter N are presented in the Table 1. It is obvious that growth of the arrivals rate
(parameter N ) causes the Gaussian approximation (28) be more accurate. If we
choose the acceptable value of the Kolmogorov distance d ≤ 0.05, we can draw
a conclusion that this approximation is applicable when value of the parameter
N (or the average number of customers at the node) is about 10 or more.
N 1 2.5 5 10 25 50 100
d 0.183 0.110 0.069 0.043 0.0026 0.016 0.011
7 Conclusions
A study of tandem of infinite-server queues with Markovian arrival process is
presented in the paper. The study is performed under an asymptotic condition
of high rate of arrival process. It is shown that the multi-dimensional joint sta-
tionary probability distribution of the customers number at the stages of tandem
can be approximated by a multi-dimensional Gaussian distribution. Parameters
of this approximation are derived in the paper. Numerical examples show that
the approximation is enough accurate when values of the arrivals rate parameter
N (or, in other words, the average number of customers at the stage) is about
10 or more.
Acknowledgments. The work is performed under the state order of the Ministry of
Education and Science of the Russian Federation (No. 1.511.2014/K).
Tandem of Infinite-Server Queues with Markovian Arrival Process 333
References
1. Grachev, V.V., Moiseev, A.N., Nazarov, A.A., Yampolsky, V.Z.: Tandem queue as
a model of the system of distributed data processing. Proceedings of TUSUR 2
(26), Part 2, 248–251 (2012). (in Russian)
2. Gnedenko, B.W., König, D. (eds.): Handbuch der Bedienungstheorie: II. Formeln
und andere Ergebnisse. Akademie-Verlag, Berlin (1984)
3. Kim, C., Dudin, A., Klimenok, V., Taramin, O.: A Tandem BM AP/G/1 →
•/M/N/0 queue with group occupation of servers at the second station. Math.
Prob. Eng. 2012, Article ID 324604 (2012)
4. Gopalan, M.N., Anantharaman, N.: Stochastic modelling of a two stage transfer
line production system with end buffer and random demand. Microelectron. Reliab.
32(1–2), 11–15 (1992)
5. Genadis, T.: The distribution of the passage time in a twostation reliable produc-
tion line: an exact analytic solution. Int. J. Qual. Reliab. Manag. 14(9), 12–25
(1997)
6. Heyman, D.P., Lucantoni, D.: Modelling multiple IP traffic streams with rate lim-
its. IEEE/ACM Trans. Netw. 11, 948–958 (2003)
7. Chakravarthy, S.R.: Markovian arrival processes. Wiley Encyclopedia of Opera-
tions Research and Management Science (2010)
8. Moiseev, A.: Asymptotic Analysis of the Queueing Network SM − (GI/∞)K . In:
Dudin, A., et al. (eds.) ITMM 2015. CCIS, vol. 564, pp. 73–84. Springer, Heidelberg
(2015)
9. Massey, W.A., Whitt, W.: Networks of infinite-server queues with nonstationary
poisson input. Queueing Syst. 13, 183–250 (1993)
10. Nazarov, A., Moiseev, A.: Calculation of the probability that a gaussian vector
falls in the hyperellipsoid with the uniform density. In: International Conference
on Application of Information and Communication Technology and Statistics in
Economy and Education (ICAICTSEE-2013), pp. 519–526. University of National
and World Economy, Sofia, Bulgaria (2014)
11. Moiseev, A.N., Nazarov, A.A.: Asymptotic analysis of a multistage queuing system
with a high-rate renewal arrival process. Optoelectron. Instrum. Data Process.
50(2), 163–171 (2014)
12. Neuts, M.F.: A versatile markovian arrival process. J. Appl. Prob. 16, 764–779
(1979)
13. Lucantoni, D.M.: New results on the single server queue with a batch markovian
arrival process. Stochast. Models 7, 1–46 (1991)
14. Nazarov, A., Moiseev, A.: Analysis of an open non-markovian GI − (GI|∞)K
queueing network with high-rate renewal arrival process. Prob. Inf. Transm. 49(2),
167–178 (2013)
15. Nazarov, A.A., Moiseeva, S.P.: Asymptotic Analysis Method in the Queueing The-
ory. NTL, Tomsk (2006). (in Russian)
16. Kolmogorov, A.: Sulla determinazione empirica di una Legge di distribuzione.
Giornale dell’ Intituto Italiano degli Attuari 4, 83–91 (1933)
A Number of Customers in the System
with Server Vacations
Abstract. We review the queueing system with one service device and
a queue with an unlimited number of waiting seats with server vaca-
tions. By using method of asymptotic analysis under large load, we found
asymptotic distribution of probabilities of a number of customers in the
system. It is shown that this probability distribution can be approxi-
mated by exponential distribution.
1 Introduction
Single-line queueing systems with server vacations are mathematical models of
telecommunication systems, which are pretty common in practice [1]. In real
systems “vacations” are considered as a temporal suspension of service either
for device other applications or for its breakdown or repair [2].
The mathematical model under study is presented in the Sect. 2. Kolmogorov
equations for the investigated processes can be found in the Sect. 3. Asymptotic
analysis of the obtained equations is performed in the Sect. 4.
2 Mathematical Model
Let’s review the queueing system with one service device and a queue with an
unlimited number of waiting seats [3]. The system receives Poisson process of
requests with intensity λ. Device operation mode consists of two consecutive
intervals. During first intervals customers are handled at the device for random
time, distributed by exponential law with parameter μ. If there are no requests
in the queue at the beginning of this interval or if the device has handled all
customers that were in the queue at that interval of time, then device is still
operating in the same mode, waiting for requests. When this interval ends, the
server goes on a vacation. During vacations, all requests that came into system
are gathered in the queue and are waiting for device to return to operating mode.
Durations of these intervals are random and determined by distribution func-
tions T1 (x) and T2 (x) respectively. We will review systems with customers pri-
ority service [4].
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 334–343, 2016.
DOI: 10.1007/978-3-319-30843-2 35
A Number of Customers in the System with Server Vacations 335
Let’s denote:
i(t) is number of customers in the system at the time t.
k(t) is device mode: 1 device is on the service, 2 device is on vacations.
z(t) is remaining time of device staying in corresponding mode.
3 Kolmogorov Equations
Let’s review three-dimensional Markov process {i(t), k(t), z(t)} and for distrib-
ution of probabilities
P1 (i, z − Δt, t + Δt) = (P1 (i, z, t) − P1 (i, Δt, t)) (1 − λΔt) (1 − μΔt)
+ P1 (i − 1, z, t)λΔt + P1 (i + 1, z, t)μΔt
+ P2 (i, Δt, t)T1 (z) + o(Δt),
P2 (i, z − Δt, t + Δt) = (P2 (i, z, t) − P2 (i, Δt, t)) (1 − λΔt) + P2 (i − 1, z, t)λΔt
+ P1 (i, Δt, t)T2 (z) + o(Δt).
∂P1 (i, z, t) ∂P1 (i, z, t) ∂P1 (i, 0, t)
= − − P1 (i, z, t) (λ + μ)
∂t ∂z ∂z
∂P1 (i, 0, t)
− P2 (i, z, t)λ + P2 (i − 1, z, t)λ + T2 (z).
∂z
We will assume that system operates in stationary mode:
⎧ ∂P (i,z) ∂P (i,0)
⎪
⎪
1
− 1∂z
⎪
⎨ ∂z ∂P2 (i,0)
+ (P1 (i − 1, z) − P1 (i, z)) λ + (P1 (i + 1, z) − P1 (i, z)) μ + ∂z T1 (z) = 0,
⎪ ∂P2∂z
⎪
(i,z)
− ∂P2∂z(i,0)
⎪
⎩
+λ (P2 (i − 1, z) − P2 (i, z)) + ∂P1∂z(i,0) T2 (z) = 0.
Let’s introduce partial characteristic functions
∞
Hk (u, z) = ejui Pk (i, z),
i=0
for which we will rewrite the direct system of Kolmogorovs differential equations
in this form:
⎧
⎪ ∂H1 (u,z)
⎨ ∂z − ∂H1∂z (u,0)
+ λ eju − 1 + μ e−ju − 1 H1 (u, z)
+ ∂H2∂z
(u,0)
T1 (z) + 1 − e−ju μP1 (0, z) = 0, (1)
⎩ ∂H2 (u,z) − ∂H2 (u,0) + λH (u, z) eju − 1 + ∂H1 (u,0) T (z) = 0.
⎪
∂z ∂z 2 ∂z 2
P1 (0, z) is probability of situation where device stays in the service mode, and
there are no customers in the system.
We will solve system (1) by using method of asymptotic analysis in conditions
of large system load [5–7].
336 A. Nazarov and S. Paul
λ = (1 − ε)Sμ, u = εw,
Fk (w, z) = Rk (z)Φ(w),
where
T1
R1 = S, R2 = 1 − S, S= ,
(T1 + T2 )
and asymptotic characteristic function Φ(w) is defined by
S
Φ(w) = ,
S − jw {S + Δ}
the value of Δ is
(2) (2)
T1 T2 T T
Δ = −R1 R2 μ + R12 R2 μ 2 + R22 R1 μ 1 , (4)
T1 + T2 2T2 2T1
where
∞
(1 − Tk (x))dx = Tk
0
(2)
is the average time of staying in the corresponding mode. Tk is the second
initial moment of the time of devices staying in mode k.
Prelimit characteristic function H(u) the number of customers i(t) can be
written as
d
H(u) = ,
d − ju
where
S
d= . (5)
S+Δ
A Number of Customers in the System with Server Vacations 337
I.e in the form of the characteristic function of exponential distribution the para-
meter d.
Fk (w, z) = Rk (z)Φ(w).
Let’s denote
∞
(1 − Tk (x))dx = Tk
0
as the average time of staying in the corresponding mode.
Then, from the one side
R1 (∞) + R2 (∞) = 1,
and from the other
∞ ∞
R1 (∞) + R2 (∞) = R (0) (1 − T1 (x))dx + (1 − T2 (x)) = dxR (0) (T1 + T2 ) .
0 0
We will obtain
1
R (0) = .
(T1 + T2 )
Then
⎧ ∞
⎪
⎪
⎨ R1 (∞) = (1 − T1 (x))dx = T1 +T2 ,
1 T1
T1 +T2
0
∞
⎪
⎪
⎩ R2 (∞) = (R2 (0) − R1 (0)T2 (x))dx = T1 +T2 .
T2
0
R1 (z) − R (0) (1 − T1 (z)) + jεw (f1 (z) − f1 (0) + f2 (0)T1 (z))
+ jwε (S − 1) μR1 (z) + O(ε2 ) = 0,
R2 (z) − R (0) (1 − T2 (z)) + jεw (f2 (z) − f2 (0) + f1 (0)T2 (z))
+ jεwSμR2 (z) + O(ε2 ) = 0.
jεw (f1 (z) − f1 (0) + f2 (0)T1 (z)) + jwε (S − 1) μR1 (z) + O(ε2 ) = 0,
jεw (f2 (z) − f2 (0) + f1 (0)T2 (z)) + jεwSμR2 (z) + O(ε2 ) = 0.
Then
⎧ ∞
⎪
⎪ {f2 (0) − f2 (0)T1 (x) + μR1 (R1 − 1) + R1 (x)μ(1 − R1 )} dx,
⎨ f1 (∞) =
0
∞
⎪
⎪
⎩ f2 (∞) = {f1 (0) − f1 (0)T2 (x) − μR1 (R1 − 1) − μR1 R2 (x)}dx.
0
+ μπ1 (∞) = 0.
Let’s assume that w = 0 and considering that
Φ(0) = 1,
we will obtain:
π1 = S, π1 = lim π1 (ε2 ).
ε→0
Then
S
Φ(w) = .
S − jw {S + (S − 1) f1 (∞) + Sf2 (∞)}
Let’s review the expression in the denominator separately
Δk = (Rk − Rk (x)) dx
0
∞
⎧ ⎫
⎨ z
⎬
1
= Rk 1− (1 − Tk (x)) dx dz
⎩ Tk ⎭
0 0
⎧ ⎫ ∞ ⎧ ⎫
⎨ z
⎬ ∞
⎨ z
⎬
1 1
= Rk 1 − (1 − Tk (x)) dx z − Rk zd 1 − (1 − Tk (x)) dx
⎩ Tk ⎭ ⎩ Tk ⎭
0 z=0 0 0
342 A. Nazarov and S. Paul
∞
1
= Rk z (1 − Tk (z)) dz
Tk
0
∞ ∞
1 z2 1 z 2
= Rk (1 − Tk (z)) d = Rk (1 − Tk (z))
Tk 2 Tk 2 z=0
0
∞ ∞
(2)
1 z2 1 z2 T
− Rk d (1 − Tk (z)) = Rk d (Tk (z)) = Rk k .
Tk 2 Tk 2 2Tk
0 0
where
∞
⎧ ⎫
⎨ 1
z
⎬
1− (1 − Tk (x)) dx dz
⎩ Tk ⎭
0 0
(2)
is the average value of the remaining time of staying in mode k. Tk is the
second initial moment of the time of devices staying in mode k. Rk (z) is written
down in this form:
⎧ ⎫
Tk ⎨ 1 ⎬
z
Rk (z) = (1 − Tk (x)) dx .
T1 + T2 ⎩ Tk ⎭
0
Then
S
Φ(w) = ,
S − jw {S + Δ}
where Δ is determined by equation (4). I.e. the function Φ(w) is a characteristic
function of the exponentially distributed random variable. Let’s denote
S
d= .
S+Δ
The theorem was proved.
By making backward substitutions, we will tend z to infinity and ε to zero,
z → ∞, ε → 0, and well get the characteristic function of a number of cus-
tomers in the system
d
H(u) = F (w, ε) = Φ(w)Rk (∞) + o(ε) = .
d − ju
k k
5 Conclusions
In this work we have researched mathematical model of the system with server
vacations. By using method of asymptotic analysis under large load, we have
found asymptotic distribution of probabilities of values of a number of customers
in the system. It is shown that this distribution is exponential with the parameter
determined by the equality (5).
Acknowledgments. The work is performed under the state order of the Ministry of
Education and Science of the Russian Federation (No. 1.511.2014/K).
A Number of Customers in the System with Server Vacations 343
References
1. Pechinkin, A.V., Sokolov, I.A.: Queueing system with an unreliable device in discrete
time. J. Inform. Appl. 5(4), 6–17 (2005). (in Russian)
2. Saksonov, E.A.: Method of the calculation of probabilities of modes for one-line
queueing systems with the server vacation. J. Automat. Tele-Mech. 1, 101–106
(1995). (in Russian)
3. Nazarov, A.A., Terpugov, A.F.: Queueing theory: educational material. NTL, Tomsk
(2004). (in Russian)
4. Nazarov A.A., Paul S.V.: Research of queueing system with the server vacation that
is controlled by T-strategy. In: Proceedings of the International Science Conference
Theory of Probabilies, Random Processes, Mathematical Statistics and Applica-
tions, pp. 202–207. Minsk (2015). ( in Russian)
5. Nazarov, A.A., Moiseeva, S.P.: Method of Asymptotic Analysis in Queueing Theory.
NTL, Tomsk (2006). (in Russian)
6. Moiseeva, E.A., Nazarov, A.A.: Research of RQ-system MMP—GI—1 by using
method of asymptotic analysis under large load. TSUs herald/messenger. Adm.
Calculating Tech. Inform. 4(25), 83–94 (2013)
7. Nazarov, A.A., Moiseev, A.N.: Analysis of an open non-Markovian GI(GI-)K queue-
ing network with high-rate renewal arrival process. Prob. Inform. Transm. 49(2),
pp. 167–178. doi:10.1134/S0032946013020063
Stationary Blocking Probability in Multi-server
Finite Queuing System with Ordered Entry
and Poisson Arrivals
1 Introduction
Queuing systems with ordered entry service discipline is a well-known type
of multi-server queueing systems in which customers are dispatched to servers
according to a selection rule known as the “ordered entry rule”. This rule implies,
that if, for example, a queueing system has N servers, numbered from 1 to N , no
queues (i.e. no waiting room) then any arriving customer who finds more than
This work was supported by the Russian Foundation for Basic Research (grants
15-07-03007, 15-07-02354, 15-07-02341).
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 344–357, 2016.
DOI: 10.1007/978-3-319-30843-2 36
Stationary Blocking Probability in Multi-server Ordered Entry Queue 345
one server idle, selects the one with the lowest number. An arriving customer
which finds all servers busy is lost. Since the seminal papers [1,2], such systems
have received significant attention from the research community. They are con-
sidered to be useful in the manufacturing processes for performance modelling
of different variations of conveyor systems (open and closed-loop) with multiple
unloading stations. Some of the early theoretical results and ideas of practical
implementations one can find, for example, in [1,3–6].
Up to nowadays queuing systems with ordered entry have already been
analysed in different setting (see [7–21]). Some of them prove to be relatively easy
to analyse and others are notoriously hard and thus approximations are used.
The main assumption in which most of queuing systems with ordered entry con-
sidered in the literature differ from one another is the assumption about the
presence of waiting rooms (queues) for the arriving customers. The analysis is
usually carried out under one of the three main assumptions: (i) there is no
waiting room for customers, (ii) there is a single queue where customers can
wait for service, (iii) each server has its own dedicated queue and any arriving
customer which finds more than one queue with free waiting room selects the
one with the lowest (server’s) number. With respect to the blocked customers
(i.e. those which find all servers busy and no free waiting room on arrival) in
the majority of cases, considered in the literature, it is assumed that the blocked
customers are lost and never return to the system. But it is also possible, as
mentioned in [20,21], for the blocked customers to recirculate i.e. each blocked
customer after some time arrives again at the system and follows the “ordered
entry rule”. Consecutive moments of such arrivals constitute the new, additional
flow of “recirculating” customers.
Note that assumptions (i)–(iii) follow directly from those problem settings,
which one usually comes across in conveyor theory. So far, most of the analytical
results for the queueing systems with ordered entry have been obtained for the
steady state under assumptions (i) and (ii). For the respective latest, though
short, review of these results, one can refer to [19]. Usually the characteristics
under study are the stationary blocking probability, the utilization of each server
and the joint stationary distribution of the number of customers in the system.
This paper is devoted to the steady-state analysis of the finite capacity
Markovian queueing system with ordered entry under assumption (iii). Such sys-
tems have not received a great deal of attention (see [9,13,15–17,20]), though are
reported to adequately model several types of real-life manufacturing systems
(see [3,4,9,10]). From the theoretical point of view, there is no general method
for their stationary analysis (except for the celebrated QBD approach) and the
presence of finite capacity queues in front of each server makes the analytical
methods used in the analysis of systems under assumptions (i) and (ii) hardly
applicable.
In [20] authors consider the system with queues at each server (with differ-
ent capacities) and possible recirculation of blocked customers. Though authors
present only the results of the thorough simulation analysis, their conclusions
contain many insightful observations concerning the interrelation of queues
346 R. Razumchik and I. Zaryadov
solution for the considered Markov chain was considered to be intractable (see
[26, Chap. 3]).
The paper is structured as follows. In the next section we give the detailed
description of the system and after that, in Sect. 3, the stationary blocking prob-
ability is being obtained at first for the 2-server system and then it is shown how
the procedure can be generalized for the N -server system (N ≥ 3). In the con-
clusion we briefly discuss the pros and cons of the proposed method and give
directions of further research.
2 System Description
Consider the N -server ordered entry queuing system (N ≥ 2), where servers are
labelled by numbers 1, 2,. . . , N without repetitions. The ith server has a buffer
of finite capacity (Ri − 1) in front of it. Customers arrive according to a Poisson
process of rate λ. The ordered entry discipline implies the following admission
rule for incoming customers. A new customer upon arrival goes to the 1st server
or, if it is busy, occupies one place in the queue in front of it. If a new customer
upon arrival sees the queue in front of the 1st server full, it goes to the 2nd server
or, if it is busy, enters its queue. Next, if upon arrival of a new customer queues
in front of the 1st and 2nd servers are full, it goes to the 3d server or, if it is busy,
to its queue and so on. Whenever newly arriving customer, sees queues in front
of each server full, it leaves the system and never comes back (i.e. recirculation
is not permitted).
In order to complete the description we specify that customers from each
queue are served according to FIFO (or LIFO, Random) service discipline and
ith server, 1 ≤ i ≤ N , serves customers for exponentially distributed times at
rate μi .
Denote by Yi (t) the number of customers in ith server and its queue at the
time t. The random process {Y (t) = (Y1 (t), Y2 (t), . . . , YN (t)), t ≥ 0} describing
the stochastic behaviour of the considered system is Markovian with continuous
time and discrete finite state space. The set of states of the process {Y (t), t ≥ 0}
is Y = {(i1 , i2 , . . . , iN ), 0 ≤ ij ≤ Rj , 1 ≤ j ≤ N }. The state (0, 0, . . . , 0)
corresponds to the empty system. The state (i1 , i2 , . . . , iN ) of the process means
that at the instant t there are i1 customers in the 1st server and its queue, i2
customers in the 2nd server and its queue etc. The process is {Y (t), t ≥ 0}
is clearly ergodic and its stationary distribution exists. Denote by {pi1 ,i2 ,...,iN ,
0 ≤ in ≤ Rn , 1 ≤ n ≤ N }, the stationary probability of the fact that there are
i1 in the 1st server and its queue, i2 customers in the 2nd server and its queue
etc. i.e.
R1
R2
P (u, v) = pij ui v j , 0 ≤ u, v ≤ 1,
i=0 j=0
R1
R2
+ μ2 u(v − 1) pi,0 u + λu
i R1 +1
v(v − u) pR1 ,j v j . (1)
i=0 j=0
The expression in the square brackets on the left hand-side of (1) is a poly-
nomial in u with roots
(λ + μ1 + μ2 )v − μ2 ∓ ((λ + μ1 + μ2 )v − μ2 )2 − 4λμ1 v 2
u1,2 (v) = u1,2 = .
2λv
Taking into consideration that we have defined the generating function
P (u, v) for the values u and v both in [0, 1], it can be seen that roots u1 and u2
are real whenever
μ2 μ2
v ∈ V = 0, √ √ ∪ √ √ ,1 . (2)
( λ + μ1 )2 + μ2 ( λ − μ1 )2 + μ2
Stationary Blocking Probability in Multi-server Ordered Entry Queue 349
R2
λuR
1 v
1 +1 R2 +1
(1 − v)pR1 ,R2 + μ1 v(u1 − 1) p0,j v j
j=0
R1
R2
+ μ2 u1 (v − 1) pi,0 ui1 + λuR
1
1 +1
v(v − u1 ) pR1 ,j v j = 0, (3)
i=0 j=0
R2
λuR
2 v
1 +1 R2 +1
(1 − v)pR1 ,R2 + μ1 v(u2 − 1) p0,j v j
j=0
R1
R2
+ μ2 u2 (v − 1) pi,0 ui2 + λuR
2
1 +1
v(v − u2 ) pR1 ,j v j = 0. (4)
i=0 j=0
Now we cancel out from these equations the term with pR1 ,R2 and then the
R2
term with j=0 pR1 ,j v j . By expressing pR1 ,R2 from (3) and substituting it into
(4) we obtain the following relation (after collecting the common terms):
R
uR − uR μ1 uR 2
2 − u1 μ2 (v − 1)
1 +1 1 +1 1 R1
2 1
− · p0,j v j+1 −
u2 − u1 λ u2 − u1 j=0
λ
R1
μ1 i uR 1 −i
− uR 1 −i
μ1 R1
R2
× pi,0 2 1
− pR1 ,j v j+1 = 0. (5)
i=0
λ u 2 − u 1 λ j=0
R2
By analogy, expressing j=0 pR1 ,j v j from (3) and substituting it into (4),
yields:
μ1 R1 +1 uR 1 +1
− uR 1 +1
μ1 uR 1
− uR 1
+ μ1 μ2 (v − 1) pi,0 2 1
= 0. (6)
i=0
λ u2 − u1
It is well-known that such fractions like (ui2 − ui1 )/(u2 − u1 ), i ≥ 1, are poly-
nomial functions of a single variable (of integer degree). Following the derivation
350 R. Razumchik and I. Zaryadov
lines from [23] one can show that the terms (ui2 − ui1 )/(u2 − u1 ), i ≥ 1, which
are present in both Eqs. (5) and (6), can be written in the form:
i−1
i−1
ui2 − ui1 μ1
= ai,j v −j , i ≥ 1, (7)
u2 − u1 λ j=0
where
⎡ ⎤
μ2
j
i−1
λ + μ1 + μ2
m
m ⎦
ai,j = (−1)j ⎣ Ci−m−1
m+1
(0) √ ,
λ + μ1 + μ2 m=j
λμ1 j
and m j is the binomial coefficient, and Cn (x) are the Gegenbauer polynomials.
m
The values of Cn (0) are equal to (see, for example, [24, p. 175]):
m
⎧
⎨0, n = 2k,
Cnm (0) = (−1) n2 Γ (m+ n2 ) (8)
⎩ , n = 2k.
( n )!Γ (m) 2
Here Γ (·) denotes gamma function. Now we can simplify (5) and (6) by
substituting in each term, where necessary, representation (7) instead of (ui2 −
ui1 )/(u2 − u1 ). Eventually by doing this one can show that (5) and (6) are indeed
polynomial functions of single variable v of integer degree. This is purely a
technical issue and thus we omit the intermediate calculations and just state the
final result. Note that after the above-mentioned substitution one has to perform
the multiplication of the sums with different upper bounds of summation. If we
assume, without the loss of generalization, that R1 and R2 are chosen such that
R2 ≥ R1 + 2, then Eqs. (5) and (6) can be rewritten in the form
1 −j
1
R
1
R1 R2 min(R−1+j,R2 )
√ v −j+1 p0,k r1 (k + j) + √ v j+1 p0,k r1 (k − j)
ρ1 j=1 ρ1 j=0
k=0 k=j
1 −1
R
R2
1 1
− (v − 1) r2 (j)v −j − √ R1 +1 pR1 ,j v j+1 = 0, (9)
ρ2 j=0
ρ1 j=0
pR1 ,R2
√ R1 (1 − v)v
R2 +1
ρ1
R2
R2
min(R1 +j+1,R2 )
+ r1 (0) p0,j v j+2
+ v j+1
p0,k q1 (k − j)
j=0 j=0 k=j
1 −j
R √ R −1
1
R1 R1
ρ1 1
+ v −j p0,k q1 (k + j + 1) − q2 (j)v −j − r2 (j)v −(j−2)
j=0
ρ2 j=0 ρ2 j=0
k=0
√ R −1
1
R1
ρ1 1
+ r2 (j)v −(j−1) + q2 (j)v −(j−1) = 0, (10)
ρ2 j=0 ρ2 j=0
Stationary Blocking Probability in Multi-server Ordered Entry Queue 351
−1−j
R1
μ2 r2 (j) + μ1 λ p0,k r1 (k + 1 + j) − μ2 r2 (j + 1) = 0, j = 0, R1 − 2, (12)
k=0
√
R1
μ1 λ
μ1 λ p0,k r1 (k) − μ2 r2 (0) − √ R1 pR1 ,0 = 0, (13)
k=0
ρ1
√
min(R1 +j,R2 )
μ1 λ
μ1 λ p0,k r1 (k − j) − √ R1 pR1 ,j = 0, j = 1, R2 . (14)
k=j
ρ1
By analogy, the second system follows from (10) by equating to zero each
coefficient at v i . For i = 2, R2 + 2 we have
1
r1 (0)p0,R2 − √ R1 pR1 ,R2 = 0, (15)
ρ1
1
√ R1 pR1 ,R2 + r1 (0)p0,R2 −1 + q1 (0)p0,R2 = 0, (16)
ρ1
min(R1 +j,R2 )
r1 (0)p0,j−2 + p0,k q1 (k − j + 1) = 0, j = 3, R2 . (17)
k=j−1
352 R. Razumchik and I. Zaryadov
min(R1 +2,R2 )
λμ1 r1 (0)p0,0 + λμ1 p0,k q1 (k − 1) − μ2 r2 (0) = 0. (18)
k=1
R1
R2 min(R1 +j,R2 )
R2
ρR
1 (1 − ρ1 )
1
pR1 ,j = ,
j=0
1 − ρR
1
1 +1
which immediately follows from the local balance principle. Such representation
for r2 (0) makes the system (15)–(18) of R2 + 1 equations have only R2 + 1
unknowns and thus it can be solved. Substitution of (19) into (18) yields
min(R1 +2,R2 )
R1
R2 min(R1 +j,R2 )
Recalling that all probabilities are usually found up to a constant, then the
substitution of p0,j = xj pR1 ,R2 (where xj are yet unknown constants) into the
previous equation yields the expression for the probability pR1 ,R2 in the form
√ R1
ρ1 (1−ρ1 )
R +1
1−ρ1 1
pR1 ,R2 = .
R1 min(R
1 +2,R2 )
R2 min(R
1 +j,R2 )
xk r1 (k) − xk q1 (k − 1) + xk r1 (k − j)
k=1 k=1 j=1 k=j
√
R2 min(R1 +j,R2 )
1 ρ1 R1 (1 − ρ1 )
p0,0 = − p0,k r1 (k − j)
R1 1 − ρR 1 +1
r1 (k) 1 j=1 k=j
k=0
R −1−j
1 −2 R1
R1
− p0,k r1 (k + 1 + j) + p0,k r1 (k) .
j=0 k=1 k=1
Once probabilities pR1 ,j , p0,j , j = 0, R2 and p0,0 are known the whole joint
stationary distribution pi,j can be calculated recursively from the system of
global balance equations.
If customer arrives at the N -server ordered entry queueing system and finds
queues in front of the 1st and 2nd servers full, it is said to have overflowed and it
tries to enter 3d server etc. Denote by τ1 , τ2 , . . . the successive moments at which
customer overflows from 2nd server and its queue. From the independence of
arriving and service process, and memoryless property of the exponential service
time distribution the time to the next overflow is independent of the past behav-
iour of the system and hence the sequence {τ1 , τ2 , . . . } forms a renewal process
with distribution function which we denote by G(t)1 . This renewal process gov-
erns the successive time instants at which new customers arrive at the 3d server
and its queue. We will find the distribution of time between overflows G(t) in
∞
terms of LST G̃(s) = 0 e−sx dG(t).
Let G̃ij (s), 0 ≤ i ≤ R1 , 0 ≤ j ≤ R2 , – be the time to the next overflow in
terms of LST, given i customers in the 1st server and its queue and j customers
in the 2nd server and its queue. Clearly G̃R1 ,R2 (s) = G̃(s). By applying first
step analysis it can be shown that G̃ij (s) satisfy the following system of linear
algebraic equations:
λ
G̃00 (s) = G̃10 (s), (20)
λ+s
1
G̃i0 (s) = μ1 G̃i−1,0 (s) + λG̃i+1,0 (s) , 1 ≤ i ≤ R1 − 1, (21)
λ + μ1 + s
1
G̃R1 ,0 (s) = μ1 G̃R1 −1,0 (s) + λG̃R1 ,1 (s) , (22)
λ + μ1 + s
1
G̃0,j (s) = μ2 G̃0,j−1 (s) + λG̃1,j (s) , 1 ≤ j ≤ R2 , (23)
λ + μ2 + s
1
Following [25], if one denotes the state space of the ordered entry queueing system,
consisting only of 1st and 2nd servers with their queues, by S = {(i, j), 0 ≤ i ≤
R1 , 0 ≤ j ≤ R2 } and adds virtual (absorbing) state k∗ , then G(t) equals the
distribution of the time until absorption of the process when the initial state is
(R1 , R2 ). In other words, G(t) can be interpreted as the first passage time distribution
from state (R1 , R2 ) into state k∗ .
354 R. Razumchik and I. Zaryadov
1
G̃i,j (s) = λG̃i+1,j (s) + μ1 G̃i−1,j (s) + μ2 G̃i,j−1 (s) ,
λ + μ1 + μ2 + s
1 ≤ i ≤ R1 − 1, 1 ≤ j ≤ R2 , (24)
1
G̃R1 ,j (s) = λG̃R1 ,j+1 (s) + μ1 G̃R1 −1,j (s) + μ2 G̃R1 ,j−1 (s) ,
λ + μ1 + μ2 + s
1 ≤ j ≤ R2 − 1, (25)
1
G̃R1 ,R2 (s) = λ + μ1 G̃R1 −1,R2 (s) + μ2 G̃R1 ,R2 −1 (s) . (26)
λ + μ1 + μ2 + s
The system (20)–(26) can be solved recursively: firstly express G̃2,0 (s)
through G̃1,0 (s) using (20) and (21), then express G̃i,0 (s) through G̃i−1,0 (s)
for each 3 ≤ i ≤ R1 from (21). After that find G̃R1 ,1 (s) (22) through G̃R1 ,0 (s),
and then obtain G̃1,1 (s) and G̃0,1 (s) from (24) and (23) and so on. The solution
can also be easily written out in the matrix form but due to the lack of space we
will not dwell on this and in the next subsection we show that one needs to be
able to solve the system (20)–(26) only up to such extent, that allows accurate
estimation of derivatives of G̃(s) at point s = 0.
As it was mentioned in the previous subsection the overflow process with dis-
tribution function G(t) and LST G̃(s) governs the successive arrival instants
of customers at the 3d server and its queue. Thus one can view the 3d server
and the queue in front of it as the queueing system with inter-arrival distri-
bution function G(t), exponential service times with rate μ3 and capacity R3
i.e. as standard GI/M/1/(R3 − 1) queue. Denote by π the probability that just
before an arrival of the customer at GI/M/1/(R3 − 1) its queue is full. Due
to the fact that the arriving at the N -server ordered entry queueing system
customer goes to the 4th server whenever on arrival it sees the queues in front
of the 1st , 2nd and 3d servers full (which happens with the probability π) and
also sees both queues in front of the 1st and 2nd server full (which happens
with the probability pR1 ,R2 ) then, once π is known, the probability pR1 ,R2 ,R3
is equal to pR1 ,R2 ,R3 = πpR1 ,R2 . The stationary probability π that the arriving
customer is lost in GI/M/1/(R3 − 1) can be calculated using the duality2 of
queue GI/M/1/(R3 − 1) to queue M/GI/1/R3 from [23], where it is shown that
1
π= ,
R3
p∗j
j=0
2
Paper [23] gives even more general results, but here we need only this one.
Stationary Blocking Probability in Multi-server Ordered Entry Queue 355
Knowing how to calculate all the probabilities pR1 ,R2 ,k , one can recursively
solve the global balance equations for the joint stationary distribution pi,j,k .
4 Conclusion
Remarkably that starting from calculation of probability pR1 ,R2 ,R3 and on, one
does not need to able to calculate the whole joint stationary distribution in order
to find the blocking probability pR1 ,R2 ,R3 ,...,RN . But this fact comes at price: at
each next step one has to calculate inter-overflow distribution (or the derivatives
of the corresponding LST) which becomes harder and harder as N grows large.
We were unable to find the general solution of the system for the N -server case
but note that for moderate values of N it can be solved using one of the many
available numerical software. One can also use the structure of the equations in
the system (20)–(26) to elaborate different approximations for the distribution
of time between overflows G(t) (see, for example, [30]). Clearly these heavily
depend on initial parameters and, for example, exponential approximation with
rate (λ + μ1 + μ2 ) sometimes is rough and sometimes is accurate (if the value
of λ greatly exceeds the value of μ1 + μ2 ). Here one must point the fact that
matrix-analytic (QBD) approach suits the purpose of obtaining joint stationary
distribution in the N -server ordered entry queue and by applying it one gets
the solution at once (although computational issues arise here also). Without
dwelling on the comparison of computational efficiency of the QBD solution
for the joint distribution pi,j with the recursive one, proposed in the paper,
we note that the latter possesses several nice properties. Firstly it uncovers the
interrelationships between the stationary probabilities of the states and secondly
356 R. Razumchik and I. Zaryadov
Acknowledgments. This work was supported by the Russian Foundation for Basic
Research (grants 15-07-03007, 15-07-02354, 15-07-02341).
References
1. Disney, R.L.: Some multichannel queueing problems with ordered entry. J. Ind.
Eng. 14, 105–108 (1963)
2. Elsayed, E.A., Proctor, C.L.: Ordered entry and random choice conveyors with
multiple Poisson Input. Int. J. Prod. Res. 15, 439–451 (1977)
3. Nawijn, W.M.: Stochastic Conveyor Systems. Ph.D. Dissertation, Technische
Hogeschool Twente, Afdeling der Werktuigbouwkunde (1983a)
4. Phillips, D.T., Skeith, R.W.: Ordered entry queueing networks with multiple ser-
vices and multiple queues. AIEE Trans. 1, 333–342 (1969)
5. Matsui, M.: Manufacturing and Service Enterprise with Risks: A Stochastic Man-
agement. Springer, New York (2008)
6. Gregory, G., Litton, C.D.: A conveyor model with exponential service times. Int.
J. Prod. Res. 13(1), 1–7 (1975)
7. Cooper, R.B.: Queues with ordered servers that work at different rates. Opsearch
13, 69–78 (1976)
8. Matsui, M., Fukuta, J.: On a multichannel queueing system with ordered entry
and heterogeneous servers. AIIE Trans. 9, 209–214 (1977)
9. Nawijn, W.M.: On a two-server finite queuing system with ordered entry and deter-
ministic arrivals. Eur. J. Oper. Res. 18, 388–395 (1984)
10. Nawijn, W.M.: A note on many-server queueing systems with ordered entry, with
an application to conveyor theory. J. Appl. Prob. 20, 144–152 (1983)
11. Pourbabai, B., Sonderman, D.: Service utilization factors in queueing loss systems
with ordered entry and heterogeneous servers. J. Appl. Prob. 23, 236–242 (1986)
12. Yao, D.D.: Convexity properties of the overflow in an ordered-entry system with
heterogeneous servers. Oper. Res. Lett. 5, 145–147 (1986)
Stationary Blocking Probability in Multi-server Ordered Entry Queue 357
13. Karlof, J.K., Jenkins, J.: The Behavior of a Multichannel Queueing System under
Three Queue Disciplines (2002). http://people.uncw.edu/karlof/publications/
jenkins.pdf
14. Jain, M., Shekhar, C.: State-dependent multi-channel queuing system with ordered
entry. Int. J. Eng. 15(3), 275–286 (2002)
15. Elsayed, E.A., Lin, B.W.: Transient behaviour of ordered-entry multichannel
queueing systems. Int. J. Prod. Res. 18, 491–501 (1980)
16. Elsayed, E.A.: Multichannel queueing systems with ordered entry and finite source.
Comput. Ops. Res. 10, 213–222 (1983)
17. Kodama, M., Fukuta, J.: On a mutlichannel queueing system with semi-ordered
entry. J. Inf. Optim. Sci. 15, 65–88 (1994)
18. Isguder, H.O., Celikoglu, C.C.: Minimizing the loss probability in GI/M/3/0 queue-
ing system with ordered entry. Sci. Res. Ess. 7, 963–968 (2012)
19. Isguder, O., Uzunoglu-Kocer, U.: Analysis of GI/M/n/n queueing system with
ordered entry and no waiting line. App. Math. Model. 38(3), 1024–1032 (2014)
20. Pritsker, A., Alan, B.: Application of multichannel queueing results to the analysis
of conveyor systems. J. Ind. Eng. 17(1), 14–21 (1966)
21. Shanthikumar, J.G., Yao, D.D.: Comparing ordered-entry queues with heteroge-
neous servers. Queueing Syst. 2, 235–244 (1987)
22. Zaryadov, I., Meykhanadzhyan, L., Milovanova, T., Razumchik, R.: On the method
of calculating the stationary distribution in the finite a two-channel system with
ordered input. Syst. Means Inf. 25(3), 44–59 (2015)
23. Miyazawa, M.: Complementary generating functions for the M X /GI/1/k and
GI/M Y /1/k queues and their application to the comparison for loss probabili-
ties. J. App. Prob. 27, 684–692 (1990)
24. Erdelyi, A., Bateman, H.: Higher transcendental functions, vol. II. Robert E.
Krieger Publishing Company, Malabar (1985)
25. van Doorn, E.A.: On the overflow process from a finite Markovian queue. Perform.
Eval. 4(4), 233–240 (1984)
26. Medhi, J.: Stochastic Models in Queueing Theory. Academic, Amsterdam (2003)
27. Razumchik, R.V.: Analysis of finite capacity queue with negative customers and
bunker for ousted customers using chebyshev and gegenbauer polynomials. Asia
Pac. J. Oper. Res. 31(4) (2014)
28. Avrachenkov, K.E., Vilchevsky, N.O., Shevljakov, G.L.: Priority queueing with
finite buffer size and randomized push-out mechanism. In: ACM International
Conference on Measurement and Modeling of Computer, San Diego, pp. 324–335
(2003)
29. Ilyashenko, A., Zayats, O., Muliukha, V., Laboshin, L.: Further investiga-
tions of the priority queuing system with preemptive priority and random-
ized push-out mechanism. In: Balandin, S., Andreev, S., Koucheryavy, Y. (eds.)
NEW2AN/ruSMART 2014. LNCS, vol. 8638, pp. 433–443. Springer, Heidelberg
(2014)
30. Whitt, W.: Approximating a point process by a renewal process: two basic methods.
Oper. Res. 30(1), 125–147 (1982)
On Polynomial Bounds of Convergence
for the Availability Factor
1 Introduction
Let us consider a restorable system, which may be either in the working state
def
during a random time ξ with a distribution function F1 (s) == P{ξ s}, or it
may be broken down and being restored by some service during another random
def
time η with a distribution function F2 (s) == P{η s}. All periods of working
and repairing are alternate and independent. The readiness coefficient (or avail-
ability factor) A(t) is defined as the probability that at time t the system is in
the working (= serviceable) state.
Often in the literature it is accepted that at initial time t = 0 the system
is serviceable and that it is in the beginning of its working period. We consider
a more general case assuming that the activity of the system may have started
earlier so that at t = 0 the system can be in one of the two states: perfect
functionality or complete failure; and further that before t = 0 the system already
spent time x in its current state.
Let us formalize the definition of readiness coefficient (availability factor).
We assume that ξi are random variables with a common distribution function
F1 (x) = P{ξi x}; likewise, ηi are random variables with a (another) common
distribution function F2 (x) = P{ηi x}; all of them are mutually independent.
Both authors are supported by the RFBR, project No 14-01-00319 A. For the first
author the article was prepared within the framework of a subsidy granted to the
HSE by the Government of the Russian Federation for the implementation of the
Global Competitiveness Program.
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 358–369, 2016.
DOI: 10.1007/978-3-319-30843-2 37
On Polynomial Bounds of Convergence for the Availability Factor 359
If at time t = 0 our system is working and its elapsed working time before
t = 0 equals x, then the residual time of this working period is a random variable
denoted by ξ (x) ; its distribution function is denoted by
(x) def 1 − F1 (x + s)
F1 (s) == P{ξ (x) s} = P{ξ s + x|ξ > x} = 1 − .
1 − F1 (x)
(x) def 1 − F2 (x + s)
F2 (s) == P{η (x) s} = P{η s + x|η > x} = 1 − .
1 − F2 (x)
def def
i−1
and t0 == ξ (x) , ti == ξ (x) + (ηj + ξj+1 ).
j=1
In the second case
def def
i
t1 == η (x) , ti == η (x) + (ξj−1 + ηj ),
j=2
and
def def def
i
t1 == 0, t1 == η (x)
+ ξ1 , ti == η (x)
+ ξ1 + (ηj + ξj ).
j=2
def
In this notation A(t) == P t ∈ [ti , ti ) .
i
It is well known that if distributions of ξ + η are non-arithmetical and
E ξ + E η < ∞, then there exists a limiting value
def Eξ
lim A(t) == A = .
t→∞ Eξ + Eη
Moreover, if E ξ n + E η n < ∞ for some n > 1, then
lim sup A(t) − Atn−1 < ∞
t→∞
However, the general theory provides neither the value of C(α), nor any
bound for it. Any knowledge of the value C(α) or its bound is rather important
in applications. Also, in the case where nothing was known earlier about such a
constant at all, even rough estimates could be useful. The goal of this paper is
to give explicit estimates to this constant.
This paper is an extended version of the conference publication [10]. The
Sect. 2 contains assumptions and notations; the Sect. 3 presents the main result;
the last Sect. 4 provides the full proof.
2.1 Assumptions
x
− λ(s) dx
We suppose that F1 (x) = 1 − e 0 , i.e., almost everywhere
F1 (s)
λ(s) = , and for some Λ > K1 > 3, K2 > 3
1 − F1 (s)
K1
Λ λ(s) when s > 0, (1)
1+s
1
F2 (s) 1 − whens > 0 ; (2)
(1 + s)K2
similarly,
b def
E ηb < == m2 (b), (5)
K2 − b
which suffices for the existence of A < ∞.
Notice that λ(s) is called intensity of failure of the recoverable system, of
course, while it is working.
On Polynomial Bounds of Convergence for the Availability Factor 361
2.2 Notations
def
1. Denote K == min(K1 , K2 ).
2. The behaviour of the system under consideration may be presented by the
random process
⎧
⎨ (1, t − ti ), if t ∈ [ti , ti );
def def
Xt = (nt , xt ) = n(Xt ) == nt , x(Xt ) == xt .
⎩
(2, t − ti ), if t ∈ [ti , ti+1 );
def
The state space of the process Xt is a set X == {{1, 2}×R+ } with a standard
σ-algebra.
def
Denote Sj == {(j, x), x ∈ R+ } ⊂ X (j = (1, 2)).
Let X0 = (n0 , x0 ).
3. Denote (here j = (1, 2)):
∞ ∞
def s k−1 (x) def s k−1
Mj (k) == k ds; Mj (k) == k
ds;
(1 + s)Kj 1−Fj (x) (1 + s + x)Kj
0 0
∞
def K1 e−Λs (a) def 1 − Fj (s + a)
κ(T ) == ds; Fj (s) == 1 − ;
1+T +s 1 − Fj (a)
0
(a) def (a) def (x) (y)
fj (s) == Fj (s) ; ϕx,y (s) == min f1 (s), f1 (s) ;
∞ s
def def
κx,y == ϕx,y (s) d s; Φx,y (s) == ϕx,y (u) du;
0 0
x,y (s) == F (x) (s) − Φx,y (s).
Φ
def
1
def Θ0 1
q == 1 − 1 − e−ΛR − κ(N R);
R (1 + N R)K1
∞
def (x )
Ψ (α, X0 ) == (2i + 4)α q i 1 + 1(n0 = 1)2α−1 M1 0 (α) + M2 (α)
i=0
(x0 ) α
+ 1(n0 = 2)M2 (α) + 2α−1 A + M2 (α)
(K1 − α)(K1 − α − 1)E ξ
(1 − A)α
+ + (i + 1)M1 (α) + i M2 (α) .
(K2 − α)(K2 − α − 1)E η
362 A. Veretennikov and G. Zverkina
3 Main Result
Theorem 1. Let K > 3 and let the conditions (1), (2) be satisfied. Then for the
process described earlier with initial state X0 = (n0 , x0 ), for every α ∈ (1, K − 1)
there exists a constant C(α, X0 ) < Ψ (α, X0 ) such that for all t 0 the following
inequality is true:
A(t) − A C(α, X0 ) .
(1 + t)α
4 Proof
4.1 Properties of the Process Xt
The process Xt defined in the Subsect. 2.2 (point 2.) is Markov. Moreover, it
possesses a strong Markov property. We skip the standard proof of both claims.
Note that trajectories of the process Xt are right continuous.
for any initial state X0 (again and always in the sequel j = (1, 2));
∞ ∞
If, further, we have an estimate E φ(τ (X0 , X0 )) C(X0 , X0 ) where φ(s) ↑
and φ(s) > 0 as s > 0, then we can use a strong Markov property and coupling
inequality: for all set M ∈ B(X )
P{Xt ∈ M} − P{Xt ∈ M} P {t τ (X0 , X0 )} = P {φ(t) φ(τ (X0 , X0 ))}.
Once the inequality (7) is established for the pair of processes, we may con-
clude that for the stationary process Xt with the initial distribution P and for
the process Xt starting from an arbitrary initial state X0 we get,
t ∈ M} = |P{Xt ∈ M} − P(M)}|
P{Xt ∈ M} − P{X
Note that since the right hand side here does not depend on M ⊆ X , this
inequality, of course, provides an estimate in total variation, that is,
0)
C(X
sup P{Xt ∈ M} − P(M)} .
M∈X φ(t)
A(t) − A C(X0 ) .
φ(t)
Due to (9) the same inequality holds true for Xt and Xt .
L
where the probability spaces Ωi,j are the copies of the described above space Ω L .
The construction of Zt is based on a sequence of stopping times tk , at which
1 n(Zt−0 ) = n Zt+0 + 1 n Zt−0 = n Zt+0 > 0,
i.e., of (random) times tk where one of the processes Zt and Zt – or both of
them – changes its first component.
Let t0 = 0 and denote
def def def def
mt == n(Zt ), mt == n(Zt ), zt == x(Zt ), zt == x(Zt ).
The sequence (tk ) will be built by induction. Assume that tk is already deter-
mined for some k and consider three cases.
Case 1. Suppose that Ztk = Ztk and mtk + mtk > 2 (that is, at
least one of
L L
the processes is in the set S2 ). Then on the probability space Ωk,1 × Ωk,2
−1
(zt )
def
we take an independent random variables θk == Fm k UΩk,1
L and
tk
On Polynomial Bounds of Convergence for the Availability Factor 365
−1
def (zt ) (zt )
θk == Fmk UΩk,2
L , they have a distribution functions Fm k (s) and
tk tk
(zt )
Fmk (s) respectively: they are residual times of stay of the processes Zt and Zt
tk
in the sets Smt and Smt correspondingly.
k k
def def
Denote θk == min(θk , θk ) and tk+1 == tk + θk . For t ∈ [tk , tk+1 ) define,
def def
Zt == (mtk , ztk + t − tk ); Zt == (mtk , ztk + t − tk );
def
Ztk+1 == 1{θk = θk } mtk − (−1)mtk , 0 + 1{θk = θk } mtk , ztk + tk+1 − tk ;
def
Ztk+1 == 1{θk = θk } mtk − (−1)mtk , 0 + 1{θk = θk } mtk , ztk + tk+1 − tk .
(10)
Case 2. Suppose now that Ztk = Ztk and mtk = mtk = 1. In this case, using
the idea of the “Lemma about three random variables” (see [9]) we construct on
L
one space Ωk,1 the pair of dependent random variables (θk , θk ) such that:
(zt ) (zt )
P {θk s} = F1(s); P {θk s} = F1 k (s);
k
∞
(zt ) (zt )
P{θk = θk } = min f1 k (s), f1 k (s) ds
0
∞
K1 e−Λs
= κzt ,zt ds = κ max ztk , ztk .
k k 1 + s + max ztk , ztk
0
(11)
It is easy to see that in this case the formulas (11) for θk and θk are true.
def
Next, we again denote tk+1 == tk +min(θk , θk ) and apply the same construc-
tion given in the formulae (10). This definition and (11) imply the inequality
P Ztk+1 = Ztk+1 κ max ztk , ztk .
366 A. Veretennikov and G. Zverkina
Case 3. Now, suppose Ztk = Ztk = (mtk , ztk ). In this case we construct random
z −1
def
variables θk = θk == Fmtkk
t
(UΩk,1
L ) (i.e., they are identical) with distribution
(zt ) L def
function Fmtkk (s) on the space Ωk,1 , and tk+1 == tk + θk . Here for t ∈ [tk , tk+1 ),
Zt = Zt = mtk , ztk + t − tk ; Ztk+1 = Ztk+1 = mtk − (−1)mtk , 0 .
This construction gives us the desired pair Zt = (Zt , Zt ), which satisfies (9)
and which is suitable for the successful coupling procedure.
Indeed, each of the processes Zt and Zt is an alternating, wherein periods
when these processes are in the sets S1 or S2 have the distribution functions
F1 (s) and F2 (s), respectively; the first period of their stay in the set Sn0 or
(x ) (x )
Sn0 has a distribution function Fn 0 (s) and Fn0 (s) – these properties are
0 0
guaranteed by the construction of processes. Moreover, for each of the processes
Zt and Zt considered separately periods of its stay in the sets S1 and S2 are
mutually independent.
Let us fix two initial values X0 ≡ Z0 = Z0 ≡ X0 . In this step of the proof we
will show the coupling inequality for the process Z = (Z , Z ); hence, the same
inequality will be established for the couple (X , X ).
First Hits to the Set. S 1 . For t > 0 denote,
def def
τ (t) == inf {s > t : Zt = (1, 0)} , τ (t) == inf {s > t : Zt = (1, 0)} .
These are the moments of the beginning of regeneration periods for the
processes Z and Z after the nonrandom t.
Denote also
def
τ Z0 , Z0 == max τ (0), τ (0) .
At τ (0) the regeneration period of the process Zt begins.
D
Its length equals θ = ξ + η where ξ and η were introduced in the Sect. 1.
After that, the behaviour of Z does not depend on the initial state Z0 (given
τ (0)). The same
can
be said about the process Z .
Let t > τ Z0 , Z0 . Then, there was at least one beginning of the regeneration
period of each of the processes Z and Z before t.
def
Denote ϑ (t) == (τ (t) − t) – the residual time of the last regeneration period
of Zt , which started before time t. From the corollary of W. Smith’s Key Renewal
Theorem (cf. [6, Lemma 1]), the following inequality holds true: if t > τ (Z0 , Z0 ),
then
On Polynomial Bounds of Convergence for the Availability Factor 367
E θ2 E (ξ + η)2
E ϑ (t)t > τ Z0 , Z0 = [= Θ0 ] . (12)
2E θ 2(E ξ + E η)
The same statement
applies to the process Zt .
Note that τ Z0 , Z0 τ (0) + τ (0), and, by virtue of Jensen’s inequality,
The number of regeneration periods of Zt before the processes Zt and Zt meet
each other according to the scheme from the step 4.6. (that is, any meeting
outside this scheme is ignored) is a random variable ν dominated by another
one with a geometric distribution with parameter p (ν itself has a more compli-
def def
cated distribution). Denote q == 1−p, and ς (Z0 , Z0 ) == inf {t > 0 : Zt = Zt }.
Obviously, ς (Z0 , Z0 ) τν .
D
Since we know the distribution of τ = τ (Z0 , Z0 ) and θ = ξ + η, we can
α
obtain an estimation of E (1 + ς (Z0 , Z0 )) for all α ∈ (1, K − 1) : by Jensen’s
inequality we get,
α
E 1 + ς (Z0 , Z0 )
∞
α
i
E 1 + τ (Z0 , Z0 ) + ξ + P{ν = i} (ξk + ηk )
i=1 k=1
∞
α
i
q i−1 E 1 + τ (0) + τ (0) + ξ + (ξk + ηk )
i=1 k=1
368 A. Veretennikov and G. Zverkina
∞
α α
q i−1 (2i + 4)α−1 1 + E τ (0) + E τ (0) + (i + 1)E ξ α + i E η α
i=1
∞
x
( )
q i−1 (2i + 4)α−1 1 + 1 (n0 = 1) 2α−1 M1 0 (α) + M2 (α)
i=1
(x ) (x )
+ 1 (n0 = 2) M2 0 (α) + 1 (n0 = 1) 2α−1 M1 0 (α) + M2 (α)
(x0 )
+ 1 (n0 = 2) M2 (α) + (i + 1)M1 (α) + iM2 (α)
∞
⎛ ∞
⎞
⎝ α (1 + s + x0 )α−1
⎠ 1 − F1 (x0 ) d x0
< d s
1 − F1 (x0 )
(1 + s + x0 ) 1
K Eξ + Eη
0 0
∞
α(1 + x0 )α−K1 α
d x0 =
(K1 − α)(E ξ + E η) (K1 − α)(K1 − α − 1)(E ξ + E η)
0
αA αA
= < ; (15)
(K1 − α)(K1 − α − 1)E ξ (K1 − α − 1)E ξ
(x )
analogously 1 (n0 = 2) M1 0 (α) P (d X0 )
X
α(1 − A) α(1 − A)
< < . (16)
(K2 − α)(K2 − α − 1)E η (K2 − α − 1)E η
On Polynomial Bounds of Convergence for the Availability Factor 369
the integral C (α, X0 , X0 ) P ( dX0 ), and this estimation give as an inequality
X
References
1. Borovkov, A.A.: Stochastic Processes in Queueing Theory. Springer-Verlag,
New York (1976)
2. Thorisson, H.: Coupling, Stationarity, and Regeneration. Springer, New York
(2000)
3. Gnedenko, B.V., Kovalenko, I.N.: Introduction to Queueing Theory. Birkhauser,
Boston (1989)
4. Kalashnikov, V.V.: Some properties of piecewise linear Markov processes. Teor.
Veroyatnost. i Primenen. 20(3), 571–583 (1975)
5. Klimov, G.P.: Probability Theory and Mathematical Statistics. Mir Publishers,
Moscow (1986)
6. Zverkina, G.A.: On some limit theorems following from Smith’s Theorem (2015).
arxiv:1509.06178
7. Lindvall, T.: Lectures on the Coupling Method. Wiley, New York (1992)
8. Griffeath, D.: A maximal coupling for Markov chains. Zeitschrift für Wahrschein-
lichkeitstheorie und Verwandte Gebiete 31(2), 95–106 (1975)
9. Veretennikov, A.: Coupling method for Markov chains under integral Doeblin type
condition. Theory Stoch. Process. 8(24(3–4)), 383–390 (2002)
10. Veretennikov, A., Zverkina, G.: On polynomial convergence rate of the availability
factor to its stationary value. In: Proceedings of the Eighteenth International Scien-
tific Conference on Distributed Computer and Communication Networks: Control
Computation, Communications (DCCN-2015), pp. 168–175. ICS RAS, Moscow
(2015)
Feedback in Infinite-Server Queuing Systems
1 Introduction
Due to development of the telecommunication and information-computing
systems, assessment of transmission capacity and of service quality is a very
important design stage. Queuing systems (QS) with feedback are adequate math-
ematic models of many real situations in which part of processed requests needs
to return to systems for repeating service.
In queuing theory there are two types of the models with feedback: models
with delayed feedback (with an orbit) [1,2]; models with instantaneous feedback
[3,4].
Article by D’Avignon G.R., Disney R.L [5] has covered queuing system that
has one server, two independent Poisson input flows which give two types of
requests, and an instantaneous feedback. Each type of requests has its own dis-
tribution for service time. Whether there is feedback (allowing for a repeating
service) or there isn’t decided depending on request type. This work has cov-
ered conditions of existence of stationary mode and conclusion about joint and
marginal distributions for the length of a queue.
Articles by I.S. Zaryadov [6,7] have covered the research of single-channel
QS with a recurrent input flow of requests, the service time for each call distrib-
uted by an exponential law, unlimited capacity storage system and the renewal
mechanism with repeating service.Request that is in the server will either leave
system with probability of p, or stay with probability of 1-p, while discarding all
other requests are rejected.
Real technical systems have limited number of servers, but in the case when
probability of request loss could be disregarded such systems could also be
approximated by infinite-server QS [8].
c Springer International Publishing Switzerland 2016
V. Vishnevsky and D. Kozyrev (Eds.): DCCN 2015, CCIS 601, pp. 370–377, 2016.
DOI: 10.1007/978-3-319-30843-2 38
Feedback in Infinite-Server Queuing Systems 371
2 Problem Statement
Let us consider with the queuing systems with infinite number of servers, Pois-
son process (M ), Markov-modulated Poisson process (M M P P ) and a renewal
process (GI) as models of arrivals.
Service time of a request is a random variable and is distributed by an expo-
nential law with parameter μ. The incoming call occupies any of free server, and
after finishing the service leaves the system with probability 1-r, or stays for
additional service with it probability r.
The object of the study is a number of requests that have come into reviewed
systems for additional service during time interval t (the flow of repeating
requests).
∂P (i, n, t)
= −λP (i, n, t) − iμP (i, n, t) + λP (i − 1, n, t)
∂t
+ iμrP (i, n − 1, t) + μ(1 − r)(1 + i)P (i + 1, n, t), (1)
Let us consider the queuing system with infinite number of servers, and Markov-
modulated Poisson arrival process (M M P P ), underlying by Markov chain with
a finite number of states k(t) = 1, 2, . . . , K, given matrix of infinitesimal char-
acteristics Q, i, j = 1, 2, . . . , K, and a matrix of conditional densities Λ. Let
us denote i(t) is a number of occupied devices at the instant t, n(t) is a number
of repeating requests that have come during the time interval t, k(t) is a state
of the Markov chain. Three-dimensional process {k(t), i(t), n(t)} is Markovian.
For probability distribution P (k, i, n, t) = P {k(t) = k, i(t) = i, n(t) = n} we can
write down Kolmogorov differential equation system for partial characteristic
function in form of differential matrix equation
∂H(u, w, t) ∂H(u, w, t)
+ jμ(rejw − 1 + (1 − r)e−ju ) = H(u, w, t)[(eju − 1)Λ + Q], (7)
∂t ∂u
where
H(u, w, t) = [H(1, u, w, t), H(2, u, w, t), ..., H(K, u, w, t)],
⎛ ⎞ ⎛ ⎞
λ1 0 · · · 0 q11 q12 · · · q1K
⎜ 0 λ2 · · · 0 ⎟ ⎜ q21 q22 · · · q2K ⎟
⎜ ⎟ ⎜ ⎟
Λ=⎜ . . . . ⎟ , Q = ⎜ .. .. . . .. ⎟ .
⎝ .. .. . . .. ⎠ ⎝ . . . . ⎠
0 0 · · · λK qK1 qK2 · · · qKK
Let us find the asymptotic characteristic function of the number of repeat-
ing requests in M M P P |M |∞ system during time interval t under condition of
increasing time. We will denote
Let us formulate the theorem, proof of which could be carried out the same
as the one of the Theorem 1.
374 S. Moiseeva and L. Zadiranova
λ = RΛE,
where
λ = R (0),
R(z) is a stationary probability distribution of values of the random process z(t).
6 Numerical Results
In order to evaluate the assessment of error of the probability distribution for the
reviewed process we have to compare the results of simulation with the results
obtained by using method of asymptotic analysis.
The domain of applicability of the results of asymptotic analysis will be
defined by means of Kolmogorov distance
Let probability of the return of a request back into the system for repeating
service is r = 0.1. Let us notice that for these parameters the rate of arrivals is
λ = 4, 73, and the parameter of Poisson distribution will be equal to λr/(1−r) =
0.525.
By increasing the value of an average service time 1/μ we can define when
the Kolmogorov distance will become acceptable.
Table 1 has covered values of Kolmogorov distances for the system at various
values of a service parameter μ.
Table 1. Error of approximation of the flow of repeating requests into the system by
the Poisson process at various values of μ
Table 2. Error of approximation of the flow of repeating requests into the system by
the Poisson flow at various values of μ
7 Conclusion
Thus, this work has covered mathematical models of M |M |∞, M M P P |M |∞
and GI|M |∞ systems with feedback. Using the method of asymptotic analysis
under condition of increasing service time, we obtain asymptotic approximation
Feedback in Infinite-Server Queuing Systems 377
of characteristic functions of the flow of repeating requests for each system. Its
proven that the flow of repeating requests could be approximated by the Poisson
process with a parameter λrt/(1 − r), where λ is a rate of the arrival process,
r is a probability of repeating service. The numerical analysis of the obtained
results is presented; the domain of applicability of approximation is defined.
Acknowledgments. This work is performed under the state order No. 1.511.2014/K
of the Ministry of Education and Science of the Russian Federation.
References
1. Pekoz, E.A., Joglekar, N.: Poisson traffic flow in a general feedback. J. Appl.
Probab. 39(3), 630–636 (2002)
2. Foley, R.D., Disney, R.L.: Queues with delayed feedback. Adv. Appl. Probab.
15(1), 162–182 (1983)
3. Neuts, F.: Matrix-Geometric Solutions in Stochastic Models: An Algorithmic App-
roach. John Hopkins University Press, Baltimore (1981)
4. Melikov, A.Z., Ponomarenko, L.A., Kuliyeva, K.H.N.: Calculation of the charac-
teristics of multichannel queuing system with pure losses and feedback. J. Autom.
Inf. Sci. 47(6), 19–29 (2015)
5. D’Avignon, G.R., Disney, R.L.: Queues with instantaneous feedback. Manag. Sci.
24(2), 168–180 (1977)
6. Zaryadov, I.S., Scherbanskaya, A.A.: Time characteristics of queuing system with
renovation and reservice. Physics 2, 61–65 (2014). (in Russian), Bulletin of PFUR.
Serires Mathematics. Information Sciences
7. Zaryadov, I.S., Korolkov, A.V., Milovanov, T.A., Sherbanskaya, A.A.: Mathemat-
ical model of calculating and analysis of characteristics of systems with general-
ized and repeating service. T-Comm Telecommun. Transp. 6(8), 16–20 (2014). (in
Russian)
8. Moiseeva, S.P., Zakhorol’naya, I.A.: Mathematical model of parallel retrial queue-
ing of multiple requests. Optoelectronics Instrum. Data Process. 47(6), 567–572
(2011)
9. Morozova, A.S., Moiseeva, S.P., Nazarov, A.A.: Research of QS with repeating
requests and uunlimited number of servers by means of a method of limit decom-
posing. Comput. Technol. 13(5), 88–92 (2005). (in Russian)
10. Leland, W.E., Willinger, W., Taqqu, M.S., Wilson, D.V.: On the self-similar nature
of ethernet traffic. ACM SIGCOMM Comput. Commun. Rev. 47(6), 202–213
(1995)
11. Klemm, A., Lindemann, C., Lohmann, M.: Modeling IP traffic using the batch
Markovian arrival process (extendend version). Perform. Eval. 54, 149–173 (2003)
12. Nazarov, A.A., Moiseeva, S.P.: Methods of Asymptotic Analysis in Queuing The-
ory. NTL, Tomsk (2006). (In Russian)
Author Index