CLOUD NETWORKING: CURRENT TRENDS,

PROBLEMS AND SOME SOLUTIONS

Ahmed Mohamed Abdelmoniem Sayed

PhD Student at CSE Department

Hong Kong University of Science and
Technology
Outline
1 Introduction

2 Data Center Networks

3 Internet Congestion Control

4 DCN Congestion Control

5 Our work & Future Works

2
 Cloud Computing Era
• Cloud computing is Internet-based computing,
whereby shared resources, software and
information are provided to computers and other
devices on-demand, like the electricity grid.

3
*source : Wikimedia Commons – Cloud Computing
 Why Cloud Computing?
• Elastic resources
– Expand and contract resources
– Pay-per-use
– Infrastructure on demand
• Multi-tenancy
– Multiple independent users
– Security and resource isolation
– Divide the cost of the (shared) infrastructure
• Simplify app deployment & management
– Common programming model across mobile, browser,
client, server, cloud

4
 Microsoft’s Cloud Platform
Finished Services
Dev Tools
& Solutions

Building Block
Services

Cloud
Infrastructure
Services

Global
Foundation
Services

*Source: http://research.microsoft.com/pubs/102318/Location-based%20service%20on%20the%20Cloud.pptx
 Virtualization

*source : Wikimedia Commons – Public Domain

• Multiple virtual machines on one physical machine

• Applications run unmodified as on real machine
• VM can migrate from one computer to another
• Each VM is typically owned by a tenant in public DC
6
 Multi-Tier Applications
• Applications consist of tasks
Front end
– Many separate components
Server
– Running on different VM machines
• Service-Level Agreements
– Per layer deadline
– Missed deadlines = Revenue Loss Aggregator

……
Aggregator Aggregator Aggregator

… …
Worker Worker Worker Worker Worker
7
Outline
1 Introduction

2 Data Center Networks

3 Internet Congestion Control

4 DCN Congestion Control

5 Our work & Future Works

8
Applications inside Data Centers
….

…. …. ….

Congestion

Front end Aggregator Workers

Server
Gmail, Bing, Dropbox, …
9
 The Need for Reconciliation
• Partition/Aggregate is the foundation for many large-scale web
services (e.g Google Search, Facebook Queries)

• Query [1KB-100KB]
- Delay-sensitive
• Short messages [100KB-1MB] - Large in number
- Few bytes amount
(Coordination, Control state)

• Large flows [>1MB] - Throughput-sensitive

(Data update, VM migration) - Few in number
- Large bytes amount

10
 Typical Sources of Performance Degradation
in Data Center Networks [4]

a) Incast : many flows go through the same port within a short interval  The
buffer space get exhausted  packets of some flows dropped  miss
deadline
b) Queue buildup : even with no packets are dropped  short flows
experience increased latency queued behind packets from the large flow
c) Buffer Pressure: when shallow buffered switch (shared memory) is used 
short flows on one port to be impacted by activity of long flows on other
ports.
11
Outline
1 Introduction

2 Data Center Networks

3 Internet Congestion Control

4 DCN Congestion Control

5 Our work & Future Works

12
 TCP Congestion Control
• Designed to address Internet congestion problem
– Window-based (AIMD) adjustment of sending rates.
– Assume packet losses  network congestion
– many variants: Tahoe, Reno, Vegas, Cubic, Westwood, ..
• Router assistance to TCP
– Random Early Detection (RED) : measures congestion based
on weighted moving average of queue length and either
drop/mark probabilistically
– Explicit Congestion Notification (ECN) : is used for conveying
congestion information to the senders
• Clean-slate approach
– eXplicit Congestion Control (XCP): Congestion Window +
Feedback (in ACKs)
 Differences Between DCN and
Internet/WAN
Characteristic Internet/WAN DCN

Latencies Milliseconds to Seconds Microseconds

Bandwidths Kilobits to Gigabits/s Gigabits to tens of Gbits/s

Causes of loss Congestion, link errors, … Congestion

Administration Distributed Central, single domain

Statistical Multiplexing Significant Minimal, 1-2 flows

dominate links

Incast Rare Frequent, due to

synchronized responses

14
Outline
1 Introduction

2 Data Center Networks

3 Internet Congestion Control

4 DCN Congestion Control

5 Our work & Future Works

15
 Data Center Transport
Requirements
1. High Burst Tolerance
– Incast due to Partition/Aggregate is common.

2. Low Latency
– Short flows, queries

3. High Throughput
– Continuous data updates, large file transfers

The challenge is to achieve these three

Conflicting Requirements

16
 Existing Solutions
1. Sender-Based :
– Mirco-seconds MinRTO [3] and DCTCP [4]
2. Receiver-Based :
– ICTCP [10] and PAC [11]
3. Switch-Assisted :
– PFabric [7] and Cutting-Payload [12]
4. Deadline-Aware :
– D3 [5] , D2TCP [8] and PDQ [9]
17
Data Center TCP (DCTCP)
Sender 1
ECN = Explicit Congestion Notification

ECN Mark (1
bit) Receiver

Sender 2

18
 Two Key Ideas
1. React in proportion to the extent of congestion, not
to its presence.
 Reduces variance in sending rates, lowering queuing
requirements.

ECN Marks TCP DCTCP

1011110111 Cut window by 50% Cut window by 40%

0000000001 Cut window by 50% Cut window by 5%

2. Mark based on instantaneous queue length.

 Fast feedback to better deal with bursts.
19
DCTCP Cluster results
Background Flows Query Flows

1 2 3

1 - Low latency for short flows.

2 - High throughput for long flows.
3 - High burst tolerance for query flows.
20
 DCTCP Summary
 Handles bursts well
 Keeps queuing delays low
 Achieves high throughput
 Based on ECN, a mechanisms already available in Silicon.
× Can not handle incast of very large number of senders
× Limited by the lower bound on window size
× Requires modification to sender and receiver TCP stack
× Fine-tuning of switch parameters
× Not suitable for public data centers

21
Outline
1 Introduction

2 Data Center Networks

3 Internet Congestion Control

4 DCN Congestion Control

5 Our work & Future Works

22
Our Work
• Simple yet efficient switch-assisted solution
• No modification to the TCP sender or receiver
stack.
• Solution that fits in regardless of TCP flavor.
• Appealing to public cloud operators.
• Incremental deployment is possible.
• IQM [12] at Globecom15
• RWNDQ [13,14] at Cloudnet15 and IPCCC15

23
TCP Flow Control is the answer
Flow Control is part of all TCP flavors
Data Data

Sender Receiver

ACK ACK

• TCP header has a Receive Window Field which is a major

part of TCP’s rate control (sending rate).
• Send Window = Min (Congestion Win, Receive Win).
• Hence, No modification is required to TCP
24
 IQM - Two Key Ideas
1. Switch port toward destination monitors connection setup rate.
 Count the number of SYN-ACKs and FINs.
 The difference represents the expected new connections.
 If expected number will overflow buffer  incast flag.

2. Set TCP receive window to 1 MSS during Incast.

 Proactively react to possible incast congestion event.
 Clear the buffer space occupied by elephants.
 Make room for the incoming incast traffic.
 Disable rewriting when incast event clears.
 Low computation and rewriting overhead.

25
IQM Algorithm
Switch side (Continuously monitor incoming SYN/FIN):
– If extra traffic > “limit”  raise incast flag.
– Set TCP RWND=1 MSS during incast epoch.
– Disable window rewriting when the queue drops back to
“Save thr”.
Switch Port
Limit Safe Thr
Data Data

ACK ACK

Sender and Receiver side (No Change):

Send Window = Min(Congestion Win, Receive Win)
26
Testbed Setup
• 12 servers: 1 master, 1 OVS physical machine, 5
senders and 5 receivers with OVS for the vPorts.
• Mice flows are Web page requests of 11.5 KB.
• Elephants flows are iperf long lived connections.
Recievers Master Senders

Bottleneck

Switch
27
Sample - Experimental Analysis
 Small Scale Testbed using Open vSwitch
 Scenario depicting 150 elephants against 30 Mice.
 Mice Goal: Low Latency and low variance.
 Elephants Goal: High and enough throughput

28
 RWNDQ - Two Key Ideas
1. Switch egress port toward destination is a receiver of the data.
 Buffer occupancy change over time
 Buffer occupancy reflects level of congestion.
 Locality of number of ongoing flow information.

2. Send explicit feedback by leveraging TCP receive window.

 Similar to XCP and ATM-ABR techniques.
 Receive window controls the sending rate.
 Feedback is less than ½ RTT away.
 Fast reaction to congestion events.
 Low computation and rewriting overhead.

29
 RWNDQ Algorithm
Switch side (Local window proportional to queue occupancy):
– Increase receive window when below the target.
– Decrease when we are above the queue target.
– Slow start to initially reach target fast.
Switch Port
Queue Target

Data Data

ACK ACK

Sender and Receiver side (No Change):

Send Window = Min(Congestion Win, Receive Win)
30
Sample - Experimental Analysis
 Small Scale Testbed using Open vSwitch
 Scenario depicting 200 elephants against 30 Mice.
 Mice Goal: Low Latency and low variance.
 Elephants Goal: High and enough throughput

31
Conclusion
• DCN congestion is a hot research topic
– Business needs and service agreements
– Quality of service (QoS)
• DCN congestion control is a necessity
– Incast is a very serious and frequent problem.
– Employing an efficient packet queueing-
scheduling to preserve small switch buffers
– Meeting deadlines either by achieving low
latency or building a deadline-aware
networking architecture.
32
Future Research Directions
• Leveraging functionalities of SDN
• Stability analysis and study.
• Handling persistent TCP connections.
• Adapting to varying initial congestion
window.
• Bandwidth allocation in Multi-tenant
datacenter with QoS constraints.

33
 THANKS!
QUESTIONS ARE WELCOMED

34
References
1. V. Jacobson. Congestion avoidance and control. ACM SIGCOMM Computer Communication
Review, 18:314-329, 1988.
2. J Dean and S Ghemawat. MapReduce : Simplied Data Processing on Large Clusters.
Communications of the ACM, 51:1-13, 2008.
3. Vijay Vasudevan, Amar Phanishayee, Hiral Shah, Elie Krevat, David G. Andersen, Gregory R.
Ganger, Garth A. Gibson, and Brian Mueller. Safe and effective fine-grained TCP retransmissions
for datacenter communication. ACM SIGCOMM Computer Communication Review, 39:303, 2009.
4. Mohammad Alizadeh, Albert Greenberg, David A. Maltz, Jitendra Padhye, Parveen Patel, Balaji
Prabhakar, Sudipta Sengupta, and Murari Sridharan. Data center TCP (DCTCP). ACM SIGCOMM
Computer Communication Review, 40:63, 2010.
5. Christo Wilson, Hitesh Ballani, Thomas Karagiannis, and Ant Rowstron. Better Never than Late:
Meeting Deadlines in Datacenter Networks. In Proc. ACM Conference on Communications
Architectures, Protocols and Applications (SIGCOMM'11), pages 50{61, 2011.
6. Theophilus Benson, Aditya Akella, and David a. Maltz. Network traffic characteristics of data
centers in the wild. In Proceedings of the 10th ACM SIGCOMM page 267, 2010.
7. Mohammad Alizadeh, Shuang Yang, Sachin Katti, Nick McKeown, Balaji Prabhakar, and Scott
Shenker. Deconstructing datacenter packet transport. Proceedings of the 11th ACM Workshop on
Hot Topics in Networks - HotNets-XI, pages 133-138, 2012.
8. Balajee Vamanan, Jahangir Hasan, and T.N. Vijaykumar. Deadline-aware data center tcp (d2tcp). In
Proceedings of the ACM SIGCOMM 2012 Conference on Applications, Technologies,
Architectures, and Protocols for Computer Communication, SIGCOMM '12, pages 115{126, New
York, NY, USA, 2012. ACM.

35
References
8. Chi-Yao Hong, Matthew Caesar, and P. Brighten Godfrey. Finishing flows quickly with
preemptive scheduling. In Proceedings of the ACM SIGCOMM 2012 conference on
Applications, technologies, architectures, and protocols for computer communication -
SIGCOMM '12, page 127, New York, New York, USA, August 2012. ACM Press.
9. Haitao Wu, Zhenqian Feng, Chuanxiong Guo, and Yongguang Zhang. ICTCP: Incast
congestion control for TCP in data-center networks. IEEE/ACM Transactions on Networking,
21:345{358, 2013.
10. Wei Bai, Kai Chen, Haitao Wu, Wuwei Lan, and Yangming Zhao. PAC: Taming TCP Incast
congestion Using Proactive ACK Control. In IEEE International Conference on Network
Protocols, 2014.
11. Peng Cheng, Fengyuan Ren, Ran Shu, and Chuang Lin. Catch the Whole Lot in an Action:
Rapid Precise Packet Loss Notication in Data Center. In Proceedings of the 11th USENIX
Symposium on Networked Systems Design and Implementation (NSDI 14), pages 17{28,
2014.
12. Ahmed M. Abdelmoniem and Brahim Bensaou, “Incast-Aware Switch-Assisted TCP
Congestion Control for Data Centers”, IEEE Global Communications Conference (Globecom
15).
13. Ahmed M. Abdelmoniem and Brahim Bensaou, “Reconciling Mice and Elephants in Data
Center Networks”, IEEE Cloud Networking Conference (Cloudnet 15).
14. Ahmed M. Abdelmoniem and Brahim Bensaou, “ Efficient Switch-Assisted Congestion
Control for Data Centers: an Implementation and Evaluation”, IEEE Performance Computing
and Communications Conference (IPCCC15).
36