Elementary Queuing Theory Notes

Richard Newman

Last modified 2/1/99

Intro

Queuing theory addresses analysis of systems that involve waiting for some service. The model usually includes one or more servers that render the service, a (possibly infinite) pool of customers, and some description of the arrival and service processes. If there is more than one queue for the server(s), then there may also be some policy regarding queue changes for the customers.

In working with queuing systems, it is easiest to analyze the system in steady state (after the system has started up and things have settled down). Time-dependent or transient analysis is more difficult and is not treated here.

Terms

The popular Kendall notation for queuing systems describes them as

A/B/c/K/m/Z systems,

where

If only the first three entries are given, then it is assumed that the source and queue capacity are both infinite, and that the queue is FCFS.

A and B may be

Customers arrive one at a time into the system, at times t_0 < t_1 < .... The differences between consecutive arrivals is the interarrival time, d_i = t_i - t_(i-1). We assume these k_i's to be independent and identically distributed random variables. Lambda is the mean interarrival time, and the expected interarrival time is E(d) = 1/lambda.

Arrivals in the system may follow any process, but in elementary queuing theory arrivals are assumed to be Poisson, that is,

Prob(d <= t) = 1 - e^(-l t),

where d is an interarrival time, t is a particular time, e is the base of the natural logarithm, and l (lambda) is the average arrival rate (in customers per unit time). So Poisson interarrival times follow an exponential distribution with mean lambda. The probability that exactly n customers will arrive in a time interval of duration t is

Prob(exactly n arrivals in time t) = [e^(-l t)(l t)^n]/n!

The expected number of arrivals in time t is

E(n) = lambda t

with variance

sigma^2_n = lambda t.

A very nice property of Poisson processes is that the sum of N independent Poisson processes is also a Poisson process, with its mean arrival rate the sum of the means.

Service times are also often assumed to be random. The mean service rate is mu, and the expected service time is E(s) = 1/mu.

Queue capacity may be infinite (approximating systems with large queues), finite (in which a customer may only wait if there is room in the queue), or zero (if all servers are busy, the customer is refused; a zero-capacity system is also known as a loss system).

Systems may have one server (most common model), or multiple servers, usually considered to be identical.

Traffic intensity is a measure of how busy a system is, and is defined as the ratio of mean service time to mean interarrival time,

u = lambda/mu.

Server utilization, the traffic intensity per server, is defined as

rho = u/c = lambda/(c mu)

for a c server system. The Law of Large Numbers indicates that this approximates the fraction of time a server is busy.

Little's Result

A simple yet useful and general equation is due to Little. If W = the mean time a customer spends in the system, and W_q is the mean time spent in the queue, L is the expected number of customers in the system and L_q is the expected number of customers in the queue, To visualize this, imagine a system in steady state. When the next customer arrives, paint him purple so you can keep track of him. When the purple customer leaves the queue to start obtaining service, then he will have waited W_q time since he arrived, on average. During that time, customers have been arriving at rate lambda cust/sec, so there are now lambda W_q customers behind our purple guy in line. Likewise, when he leaves the server, there will be lambda W customers in the system.

M/M/1 Systems

In general, these systems in which the state may be characterized by the number of customers in the system, and which can change state only by incrementing (birth) or decrementing (death) that number, are called Birth-Death systems. They are a special form of Markov systems, which are characterized by a set of mutually exclusive and collectively exhaustive discrete states, in which the current state and transition probabilities characterize the future behavior of the system. In other words, the system has no memory other than the state it is in (it does not matter how the system got into that state). Birth-Death systems are a special case in that the transition probabilities are zero except for adjacent states.

A state diagram for such systems appears as a linear sequence of states, labeled with the number of customers in that state 0, 1, 2, ..., i, ... and with arcs between states with consecutive numbers (i and i+1). The arc from state i to i+1 is labeled lambda (for mean arrival rate) and that from i+1 to i is labeled mu (for mean service rate). Let the probability that a particular state i is occupied be p_i. Then the expected rate at which transitions are made from state i to state i+1 is lambda p_i, while the expected rate at which transitions are made from state i+1 to state i is mu p_(i+1). At steady state, these must be equal,

mu p_(i+1) = lambda p_i, i=0,1,2,...

So

p_(i+1) = (lambda/mu) p_i = (lambda/mu)^(i+1) p_0, i=0,1,2,...

Since the sum of probabilities of existence over all states must be 1, we have

Sum [i=0 to inf] p_i = 1 = p_0 + rho p_0 + rho^2 p_0 + rho^3 p_0 + ...

= Sum [i=1 to inf] rho^i p_0 = p_0 Sum [i=1 to inf] rho^i

where rho = lambda/mu. Thus, by telescoping the series,

p_0 = (1 - rho),

which represents the probability that there are no customers in the system. The probability that there are one or more customers in the system, and thus that a new arrival must wait, is simply 1 - p_0 = rho.

To compute the mean number of customers in the system, we take the weighted sum

L = Sum [i=0 to inf] i p_i = Sum [i=1 to inf] i p_i = p_0 Sum [i=1 to inf] i rho^i

Again, by telescoping the series, we obtain

L = rho/(1 - rho)

By applying Little's result, we get the mean waiting time

W = L/lambda = E(s)/(1 - rho) = 1/(mu - lambda),

recalling that rho = lambda/mu and E(s) = 1/mu.

General formulas for M/M/1 systems are as follows:

M/M/c Systems

General formulas for M/M/c systems are as follows:
This document is copyright 1999 by Richard E. Newman.
Send comments to
nemo@cise.ufl.edu