Traffic theory an introduction

Italo TOBIA

Training for TIM Celular - Rio de Janeiro

Index
Traffic models and fundamentals

Service systems Traffic intensity Circuit utilization Pure loss systems Example Pure delay systems Example

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

The goal of this short presentation is to provide the basic concepts, results and tools which can be useful in analyzing and design telecommunication systems from a probabilistic point of view. Detailed models (such as stchastic processes, queue networks, analytical demonstrations, etc..) are not considered here because they arent essentials for the purpose of this course.

General model
Service system 1 Traffic sources 1 2 ... S K ... 1 2

Queue N Servers

S, K, N specify the service system structure

Telecommunication networks Traffic theory Italo TOBIA
Training for TIM Celular Rio de Janeiro

The classical theory studies the service systems in their statistical equilibrium point (probabilities are stationary). This is due to reasons: - Analytical models are simplified (linear algebric equations versus linear differential equations) - Time constants for the processes are very small (practically they are related to 1/( + ) ) In the stationary state we have only random variables instead of stochastic processes; we do not need to work in the time domain because the time dimension disappears.

Service system characteristics

A service system is defined by: The number of sources - N; Incoming requests flow; Queue length K; Queue scheduling system; N servers; Service time. We can have different kind of systems, depending on the relationship among N, K, S: If K = 0, we have a pure loss system; If S N + K we have a pure delay system; If S > N + K we have a hybrid system with delay and loss.

Service time is the mean time required to serve a generic request (e.g. the holding time for telephone systems)
Telecommunication networks Traffic theory Italo TOBIA
Training for TIM Celular Rio de Janeiro

Interesting random variables used in traffic calculations: X = number of customers in the system (queue + service) L = number of queued requests Y = number of busy servers W = waiting time T = total delay H = service time X=L+Y and T=W+H Our interest is focused as follows: Loss systems: Time congestion Call congestion Carried traffic Delay systems: Statistical distribution for T, W Delay probability Queuing probability Delay-loss systems All of the preceding items The average values for the performance parameters in the stationary solution, never depend on the scheduling scheme, except when the service time is related to the scheduling policy (e.g Shortest Job First or Longest Job First).

Calls during time

12 10 8 6 4 2 0
0 3 6 9 12 15 18 21 24
Traffic theory Italo TOBIA

Calls during time

Telecommunication networks

Training for TIM Celular Rio de Janeiro

The number of calls during time represented is a possible representation of the stochastic process Q(t). We observe that in a generic time interval the number of calls in the system can increase or decrease by 1.

Traffic intensity
The traffic measurement unit is the Erlang. The Erlang is the traffic intensity offered to / carried by 1 resource in 1 hour observing time in 1 hour time interval. As for telecommunication networks and systems, there is an international agree which defines this reference time interval as busy hour. If we have a N channels bundle and there is NO LOSS (that means all request can be accepted an carried on), the offered traffic is equal to the developed traffic; its given by:

A=

Ncalls Tm * 3600

Ncalls is the number of the hour calls Tm is the average holding time (in seconds)

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

Example. If we consider one subscriber that originates 3 calls per hour (during the busy hour), with 3 minutes service time (call duration), the traffic that has been offered to the system is equal to: Ao = 3 * 3/60 = 0.15 Erl Conceptually the offered traffic represents the carried traffic only when the system does not discharge any call, that is the system has infinite servers (circuits). The carried traffic represents the mean busy server number; so, the maximum carried traffic for a system with N servers, is equal to N [Erl].

Traffic balance
Statistical equilibrium relation: Offered traffic = Carried traffic + Lost traffic because Offered request = Carried requests + Lost requests We can measure carried traffic We can estimate lost traffic We can partially measure offered traffic

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

We speak about offered traffic with reference to the number of active sources. We speak about carried traffic with reference to the number of busy servers. If we say that As is the traffic carried by N servers, we mean that: -There are As busy servers in the average -As is the percentage of resources occupancy -As servers carry As Erlang if they are always busy

Carried traffic, spill-over traffic

Offered traffic

Number of servers

Spill-over traffic

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

Note that the overflow traffic has more dispersion with respect to fresh traffic and it presents several peaks. So, spill-over traffic cannot be handled in the same way as poissonian traffic (fresh traffic). Several methods could be analysed to make detailed calculations [see last page for further readings].

Circuit utilization
It is the mean traffic carried by each circuit ( ) In computer systems there is a similar parameter that is related to the CPU utilization percentage. Circuit utilization mainly depends on the resource assigning policy. As for example, if we have a bundle, and for each request a channel must be found with a fixed hunting scheme (always from the 1st to the last channel), the circuit utilization willl be decreasing from the first channel to the last. Circuit utilization will be the same for all the circuits in a bundle if the scheduling algorithm choose randomly a free line to be assigned to the generic requesting call. Otherwise, in a sequential hunting scheduling scheme, first resources absorb more traffic than the last ones. In a N circuits bundle, the mean circuit utilization is given by:

= As /N
Telecommunication networks Traffic theory Italo TOBIA
Training for TIM Celular Rio de Janeiro

The maximum circuit utilization is equal to 1, that is for a system with 1 server always busy during the busy hour. The circuit utilization is also referenced as normalized throughput.

Kendall notation
A1/A2/N/K/S Sources number Queue length Servers number Service time distribution Interarrival time distribution
A1 and A2 may be: M = Markovian or D = Deterministic or G = General etc

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

10

According to the Kendall formalism, when we do not specify the number of sources, neither the queue length, we implicitly mean that sources are infinite and the queue dept is infinite too. M/M/10 can model the behaviour of a switching system with 10 outgoing trunks, infinite local subscribers.

Hypothesis for teletraffic calculations

Requests are independent among each other Interarrival distribution time is exponential (Markovian, (M/...) Service time is exponentially distributed too (M/M/...) Scheduling is FIFO The system is analysed in the stationary state (that is, each state probability becomes constant)

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

11

The mathematical models most widely used to represent telephone traffic make the following assumptions: It is very rare for two or more calls to arrive at the same very small time interval The probability a call arrives in a given time interval is proportional to the length of that interval The probability that a call arrives at a given instant does not depend on the time neither on the past history. This assumptions are very close to reality when the number of sources (subscribers) is sufficiently large. However, some exceptions occur and must be taken into account: - Repeated calls (lost attempts are offered to the system again until they can be processed) - Internet access (long duration)

M/M/N/0 model (pure loss model)

Exponential distribution for interarrival time Exponential distribution for service time (the longer is the call, the smaller is its probability) So, we have a Poisson process (birth/death process)

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

12

A Markovian process by definition allows only transitions between adjacent states. As an example, we can model a system M/M/1 by means of a state transition diagram. In the 0 state, the server is free In the 1 state, the server is busy P {0 -> 1} = * t (transition 0 -> 1 probability) P {1 -> 0} = * t (transition 1 -> 0 probability) P {0 -> 0} = 1 - * t (no transition from state 0 probability) P {1 -> 1} = 1 - * t (no transition from state 1 probability)

1 birth 0 1 death 1

Results from M/M/N/0 model

B is the loss probability, or congestion probability N is the number of servers (circuits or channels) A is the offered load (given by /, with =average arrival rate, = inverse of the service time or the service rate) We have the Erlang B formula:

B (N, A) =

AN N!
K=0

AK K!
Traffic theory Italo TOBIA

Note: remind that K ! = 1*2*3*.*(K-1)*K and, by definition 0! = 1)

Telecommunication networks
Training for TIM Celular Rio de Janeiro

13

A represents the traffic carried by an infinite trunk group without loss. The lost traffic is given by: A*B while the carried traffic is A*(1-B) The traffic lost by a trunk group that is offered to another trunk group is known as overflow or spill-over traffic. The nature of this kind of traffic is not Poissonian and cannot be handled same as fresh traffic. Recursive calculation for B(N,A): B(N,A) = [A*B(N-1, A)] / [N + A*B(N-1, A)] Traffic decomposition: A = A*(1-B) + A*B

Offered traffic

Carried traffic

Lost traffic

Inverse Erlang B function

B = 0,01
We fix B and evaluate N as a function of the offered traffic A (that is what we need in a typical design problem)

35 30 25 20 N 15 10 5 0 4 8 12 16
Italo TOBIA

20 A
Training for TIM Celular Rio de Janeiro

Telecommunication networks

Traffic theory

14

B(N,A) represents the probability that a call request experiences congestion. Erlang B formula is very general because it is invariant with respect to the server hunting scheme. The loss probability mainly depends on the birth/death process structure and parameters.

An example for loss system

Erlang loss formula can be used: -In the design process - We estimate the offered traffic A - We know B - We can evaluate N with the inverse loss formula -In the performance evaluation process we observe the system - N is known - We can measure the carried traffic As - We evaluate B with the loss formula, solving the non linear equation: As = A*[1-B(N,A)] With respect to A; we then evaluate B = (A- As) / A or B=B(N,A)

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

15

Your calculation here

Loss system design example

Outgoing traffic A1 Incoming traffic A2

PSTN
PABX N voice circuits

Problem data: 200 PABX subscribers B =1% During the busy hour: 3 outgoing calls / 3 minutes holding time for each subscriber 3 incoming calls /4 minutes holding time for each subscriber Evaluate N, that is the number of bidirectional circuits at the interconnecting point
Traffic theory Italo TOBIA

Telecommunication networks

Training for TIM Celular Rio de Janeiro

16

As the offered load we must consider the total A1 + A2 traffic, because the bundle is bidirectional. A1 = 3 * 200 * 3 / 60 = 30 Erl A2 = 3 * 200 * 4 / 60 = 40 Erl A = A1 + A2 = 70 Erl By solving the equation: B(N, 70) 0,01 we find N 85 As a consequence, we must have 3 E1 links between the PABX and the PSTN By dimensioning the interconnection as 2 independent unidirectional links, we would have: A1 = 30 Erl, B 0,01 hence N1 42 (2 E1 links) A2 = 40 Erl, B 0,01 hence N2 53 (2 E1 links) N = N1 + N2 = 95 a more expensive solution (cards, lines, line terminals, etc..)

M/M/N model (pure delay)

The delay probability, that is, the probability a request finds the N servers busy, is given by the so called Erlang C formula

N * AN N ! * (N - A) C (N, A) = N-1 AK N * AN + N ! * (N - A) K = 0 K !

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

17

In tne Erlang delay formula A is the traffic intensity, given by /.

Waiting time distribution

The probability that queueing delay is less or equal ta, is given by the Molina formula: P { waiting_time ta } = 1 - C (N, A) *

(N - A) * ta ts

ts being the processing time The average waiting time in the queue is:

tm =

C (N, A) * ts N-A
Traffic theory Italo TOBIA
Training for TIM Celular Rio de Janeiro

Telecommunication networks

18

You need custom algorithms to evaluate C(N,A) developed with Excel macros, matlab routines or other programming languages routines. You can search over the Internet several sites containing C written useful functions to evaluate B and C Erlang formulas.

An example for delay systems

A packet multiplexer with the following operating parameters and data: C= physical capacity for the outgoing channel = 50 packets/s Packet length exponentially distributed with mean 100 byte 95% of the packets must be in the queue for a time less than or equal to 0.1 s Evaluate the queue length

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

19

The offered load can be evaluated imposing the limitation on the waiting time in the queue:

P { waiting_time tw } = 1 - C (N, A) *

(N - A) * e ts

tw

With P 0.95; ts the service time which is equal to A/; tw = 0.1 s; one server, that is N=1. Remember that C(1,A)=A The equation above can be solved numerically evaluating A which is 0.66 and then: C = * L / A = 60.64 kbps L being the mean number of bit for a packet

Designing the number of operators in a call center

1 min service time 600 calls/hour P{waiting_time 2 min} 0.95 Find the number of the operators - N

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

20

A = ts * = 1*600/60 = 10 Erl Applying the Molina formula, we have: C(N,10) * exp[-(N-10)*2/1] (1-0.95) As for the system stability, we must have N > A, that is N>10 After 2 attempts, we find that N=12 satisfies the problem requirements

For more details

J.D.C. LITTLE: A proof of the queueing formula L=W Operation research, 9, 383 387, 1961 A. PAPOULIS: Probability, random variables and stochastic processes Mc Graw Hill 1964 M. SCHWARTZ: Telecommunication Networks, protocols, modeling and analysis Addison Wesley 1987 R. SYSKY: Introduction to congestion theory in telephone systems Oliver and Boyd 1960

Telecommunication networks

Traffic theory Italo TOBIA

Training for TIM Celular Rio de Janeiro

21