J.

Virtamo

38.3141 Teletrac Theory / Overow trac

OVERFLOW TRAFFIC Overow trac in a loss system

Consider a circuit switched network which works as a loss system (blocked calls are cleared). System is oered Poissonian trac with intensity . The holding times are exponentially distributed with parameter . The oered trac intensity a = /. The system consists of two parts: - primary system: m1 servers - secondary system: m2 servers The arriving trac is rst oered to the primary system. If all servers of the primary group are occupied the call is directed to the secondary group, the trac directed from the primary group to the secondary group is called overow trac; denote its intensity by . If all servers of the secondary group are also occupied, the arriving call is blocked.
sekundriryhm m2 palvelinta a ylivuotoliikenteen intensiteetti primriryhm m1 palvelinta a = l/m tarjotun liikenteen intensiteetti

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Trac process of the overow trac

N2(t)

The primary and secondary groups together constitute a m1 + m2 server loss system (Erlangs system; M/M/m/m queue, where m = m1 + m2 ). The primary system alone forms an m1 server loss system (m = m1 ). The division of the capacity m1 + m2 of the system into two parts does not aect the overall behaviour of the system (the blocking probability is the same as in an m1 +m2 server system). One can think the trunks of the system to be numbered. An arriving call is carried by the free trunk with the lowest number.

m2

N1(t) m1

N1 truncated (m1 ) Poisson distribution N = N1 + N2 truncated (m1 + m2 ) Poisson distr. N2 = N N1 occupancy of the secondary group

Note. Though both N and N1 are separately insensitive (independent of the holding time distribution), the distribution of N2 is not insensitive.

Overow trac is generated only when the primary group is in blocking state. The blocking periods of the primary group form time windows, during which the arriving trac is directed to the secondary group. The arrival process of the overow trac is so called interrupted Poisson process (IPP).

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Blocking of the overow trac Intensity of the overow trac (trac blocked in the primary group) = a E(m1 , a) where E(m, a) is the Erlang B-formula. Ultimately blocked trac = a (trac blocked in the secondary group) a = a E(m1 + m2 , a)
a al m2 ac m1+m2 a m1 ac al a

Call blocking of the overow trac B2 = aE(m1 + m2 , a) aE(m1 , a) B2 = E(m1 + m2, a) E(m1 , a) One can show that B2 > E(m2 , )

The blocking experienced by the overow trac is greater than the blocking that would be experienced by Poisson trac with the same intensity in a secondary group of m2 trunks. This is due to the fact that the IPP process is more bursty than the ordinary Poisson process.

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Blocking of the overow trac (continued) Example Assume that the intensity of the oered trac is a = 5 erl. The sizes of the primary and secondary groups: m1 = 5 and m2 = 1

overow trac nally blocked trac

= 5 E(5, 5) = 5 0.285 = 1.42 a = 5 E(6, 5) = 5 0.192 = 0.96 a E(6, 5) = = 0.67 E(5, 5) E(1, 1.42) = 0.59 B2 =

blocking of overow trac blocking if the trac were Poissonian

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Peakedness of the overow trac Sometimes it is useful to characterize a trac process by telling what would be the occupancy distribution if the trac were oered to a trunk group of innite capacity. To characterize the overow trac in this way, assume now the secondary group is innite, m2 = . Then all the overow trac will be carried in the secondary group and it holds E[N2] = aE(m1 , a) = Also the variance of N2 can be calculated in this case. The derivation is rather involved. The result is known as the Riordan formula: V[N2] = 1 + a m1 + 1 a +

The variance to mean ratio of the occupancy is called the peakedness factor. (In the case of a Poisson arrival process the occupancy distribution is Poisson(a) distribution with the peakedness factor 1.) z= V[N2] a = 1+ E[N2 ] m1 + 1 a + where = aE(m1 , a)

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Peakedness of the overow trac (continued) The peakedness factor is a function of m1 and a, z = z(m1 , a). When a is held xed and m1 is increased - rst, for small m1 , z 1 (all the trac ows over) - then z attains a maximum (when m1 a) - nally, for very large m1 , z 1 (the overow events are rare singular events)
7 6 peakedness z 5 a=50 a=20 a=10 a=5 4 3 2 1 0 2 5 10 20 50 100 200 500 1000 number of servers m a=100

J. Virtamo

38.3141 Teletrac Theory / Overow trac

An example of overow trac The gure on the right shows the oered trac in a period of ve days a The gures below show the same trac in a primary group of 60 trunks and the overow trac. The trac in the primary group is smoother (truncated) than the oered trac, whereas the overow trac is more peaky than the oered trac.
a

From: A. Myskja, Telektronikk 2/3 (1995).

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Haywards approximation
Haywards approximation provides an approximate way to calculate the blocking probability for non-Poissonian trac (e.g. overow trac). The starting point is the observation that for the occupancy N induced by Poissonian trac in an innite server system the following relations hold: E[N] = a V[N] = a N Poisson(a)

For non-Poissonian trac, in contrast, we generally have V[N] = E[N]. The Hayward approximation tries to describe non-Poissonian trac by equivalent Poisson trac and then apply Erlangs formula for the blocking probability. The idea is to consider the behaviour of the occupied capacity R instead of the occupancy N. Let

c = the bandwidth (number of trunks) required of a single connection R = N c = the bandwidth occupied in the occupancy state N

J. Virtamo

38.3141 Teletrac Theory / Overow trac

Haywards approximation (continued)

For Poissonian trac holds

Consider a non-Poissonian source with known mean and variance of occupation

E[R] = E[c N] = c a V[R] = V[c N] = c2 a

V[R] =c E[R]

This is now replaced by Poissonian trac R a = intensity of the trac c = bandwidth requirement of a single connection so that the mean and variance of the occupied capacity are the same for the original nonPoisson trac and the model Poisson trac: E[R] = E[R ], V[R] = V[R ]

E[N] = a V[N] = v

E[R] = c a V[R] = c2 v

R' c c'

The point is that for the equivalent Poisson trac also the bandwidth required by a single connection is taken as free parameter. A single connection of the equivalent trac may thus require e.g. 1.6 trunks.

J. Virtamo

38.3141 Teletrac Theory / Overow trac

10

Haywards approximation (continued)

The tting of the two rst moments leads to the conditions for a and c

a c = ac a c 2 = v c2

a2 a = v v c = c a

equivalent intensity equivalent bandwidth

The size of the system is modied correspondingly. If the original system has m trunks, i.e. a capacity of m c bandwidth units, then it can accommodate m c/c equivalent connections. Thus the equivalent system has m trunks: mc a m = =m c v Now the blocking probability is approximated by that of the equivalent Poissonian trac: a a2 m a B E(m , a ) = E(m , ) = E( , ) v v z z Haywards approximation where z = v/a

The load per the server is the same as before: (a/z)/(m/z) = a/m. When z > 1, the system becomes smaller blocking increases.

J. Virtamo

38.3141 Teletrac Theory / Overow trac

11

Haywards approximation (continued)

The non-Poissonian trac may originate from several independent sources. If for each source the mean ai and the variance vi of the occupancy are known, then the corresponding parameters for the aggregate stream are

a =
i

ai vi
i

v =

Haywards approximation then gives the approximate total blocking probability of the aggregate stream in a system with m trunks, m a a a2 B E(m , ) = E( , ) v v z z where z = v/a

How the blocking is distributed among dierent sources remains undened.

J. Virtamo

38.3141 Teletrac Theory / Overow trac

12

The ERT method (Equivalent Random Theory)

The ERT method is also known as the Wilkinson method. This provides another approximate method to calculate the blocking probability for nonPoissonian trac. The oered trac is characterized by

In the case of several independent components we have a=

i

a = trac intensity = mean occupancy in an innite system v = the variance of the occupancy in an innite system ai , v=
i
a, v a* 1 2 m*

vi

a and m are determined such that the overow trac in the ctitious channel has the intensity a and variance v.

The idea of the ERT method is to think that the trafc (a, v) is obtained as overow trac from a ctitious channel a = oered trac = number of servers (trunks) m in the ctitious system

J. Virtamo

38.3141 Teletrac Theory / Overow trac

13

The ERT method (continued) Moment matching conditions are (the variance is given by the Riordan formula)

a = a E(m , a)

a v = a 1a+ m + 1 a + a

a, m

If the channel to which the nonPoissonian trac is oered has m trunks, the intensity a of the ultimately overown trac can be calculated a = a E(m + m, a ) An estimate for the trac blocking is correspondingly B= a a

a* 1 2 m*

a, v 1 2 m

al

J. Virtamo

38.3141 Teletrac Theory / Overow trac

14

The ERT method (continued) The solution to this pair of equations has to be found numerically. An additional diculty is that, in general, there is no solution for an integer m . One has either to make further approximations or to extend the denition of Erlangs blocking formula for real valued trunk numbers (which, of course, are unrealistic). Such an extension can indeed be made: am ea E(m, a) = (m + 1, a) where the denominator is the incomplete gamma function (m + 1, a) =
m t t e dt a

By partial integration one can show for integer values of m that (m + 1, a) =

m t t e dt a

= m!e

a am 1 + + + 1! m!

whereby the formula reduces to the familiar form E(m, a) = am m!

a am 1 + + + 1! m!

J. Virtamo

38.3141 Teletrac Theory / Overow trac

15

The ERT method (continued) For the solution one can also use approximation formulae, e.g. that suggested by Rapp (1964),

The approximation is not accurate if a is small and z is large.

a = v + 3z(z 1) a(a + z) m = a1 a+z1

where z = v/a (this is an exact relation; approximation is in a)

J. Virtamo

38.3141 Teletrac Theory / Overow trac

16

The ERT method (continued) The parameters of the ERT method The parameters of the model system, m (lower curve) and a (upper curve) solved from the pair of equations are shown in the gure on the right, when the oered trac has a = 10 and the peakedness z = v/a varies from z = 1 . . . 2. In the solution we have used the exact extension of Erlangs function for real number of trunks. Comparison of the Hayward and ERT methods The gure shows the blocking probability for trac with a = 10 and z = 1 . . . 2 oered to a m = 15 trunk system. The calculations have been done both with the Hayward method and the ERT method. In this case the results are very close to each other. One cannot say which one is more correct. The blocking of the trac, in reality, depends not only on the parameters a and v = z a (and m). Based solely on these parameters the correct result in not known.
Hayward (alempi), ERT (ylempi); a=10, m=15 ja z=1...2 0.1 0.08 0.06 esto 0.04 0.02 0 1.2 1.4 1.6 1.8 2 piikikkyys z

ERT parameters with a=10 and z=1...2 25 20 m* and a* 15 10 5 0 1.2 1.4 1.6 1.8 2 peakedness z

J. Virtamo

38.3141 Teletrac Theory / Overow trac

17

The ERT method (summary)

todellinen jrjestelm mallijrjestelm

a1, v1

a2, v2

aK, vK

a, v

m1

m2

mK

m*

A1

A2

AK

a*

The aggregate overow trac is handled as if it originated from a single channel. Find a and m such that the overow trac of the model system has the right a and v, a = a i , v = vi .
i i

aE(m + m, a) aE(m + m, a) The blocking in the overow channel is ; total blocking a Ai

i