4.10.1 Important Property of M/M/m Queueing Systems

        For M/M/m queueing systems with infinite queue capacity, we now state a property that often plays an important simplifying role in the analysis of queueing networks. It is sometimes referred to as the equivalence property for M/M/m systems.

Let the arrival process at a M/M/m queueing system with infinite queue capacity have parameter lamda.gif (291 bytes). Then, under steady-state conditions  (i.e., for lamda.gif (291 bytes) < mmu.gif (189 bytes)), the departure process from the queueing system is also Poisson with parameter lamda.gif (291 bytes).

        The proof of this property is quite involved. However, it is relatively straightforward for the special case M/M/1 (see Problem 4.10).

        The implications of the property for the analysis of queueing networks are quite obvious. If some "component" (facility) of a queueing network can be modeled as a M/M/m system with infinite capacity, the "output" of this "component" is also Poisson with parameter lamda.gif (291 bytes). That is, users will leave this specific facility according to a probability distribution identical to the probability distribution for the arrival of users at the facility.

        Thus, if the served users of the M/M/m facility are subsequently routed to another facility, the arrival process to this other facility is a Poisson process. In fact, it is clear that if

1.    the queueing network consists of, say, K facilities in series (Figure 4.17), each of which     
       contains m1, m2, . . . , mk (mi = 1, 2, 3, ...) identical servers with negative exponential 
       service and rates mu.gif (189 bytes)i, (i =1, 2, . . . , K);

2.    there is infinite queueing space between successive facilities; and

3.    the arrival process for the first facility (facility 1 in Figure 4.17) is Poisson with rate lamda.gif (291 bytes);

then, under steady-state conditions, the queueing network can be analyzed as K independent M/M/m queues and the results of Section 4.6 are directly

fig4.17.gif (25576 bytes)

FIGURE 4.17 Poisson input at rate lamda.gif (291 bytes) at queueing system 1 will result in a Poisson input at rate lamda.gif (291 bytes) for all four queueing systems.

applicable. The condition for steady state in this case is that lamda.gif (291 bytes) < mimu.gif (189 bytes)i, for all i(i = 1, 2, . . . , K).
Three additional notes:

1.  It is not necessary that all users proceed in series from one facility to the next. For instance, in a "feedforward" network (i.e. a network without any feedback loops) of the type shown in Figure 4.18, departing users from a facility may be divided among N subsequent facilities according to the probabilities pj (j = 1, 2, . . . , N; sumnj1.gif (1222 bytes) pj=1. Here pj represents the probability that a user departing the facility will go to facility j, with the routing of each user determined independently. Since the streams of events that result from the subdivision (in the way described above) or from the mixing together of Poisson streams are also Poisson, the M/M/m analysis can also be used in such, more complicated cases.14 Thus, as an example for the network shown in Figure 4.18, all flows between facilities are Poisson as long as the arrival processes represented by lamda.gif (291 bytes)1, lamda.gif (291 bytes)2, and lamda.gif (291 bytes)3, are Poisson and as long as the assumptions in the equivalence property hold for each queueing system separately.

2.    The equivalence property does not hold for the case of M/M/m

fig4.18.gif (39605 bytes)

queueing systems with finite queue capacity. It is simple using a state-transition diagram to argue intuitively why this is so. Queueing networks composed of M/M/m queues of this type are often analyzed using the state-transition diagram approach that we shall describe below.

3.    Unfortunately, the "negative" statement of the equivalence property is also true: it can be shown [GROS 74] that the only type of service distribution which results in a Poisson output, if the input to the queueing system is Poisson, is the negative exponential distribution. Thus, a necessary condition for the output (the departure process) of a M/G/m queueing system to be Poisson is that "G = M" (i.e., that the service time distribution be negative exponential). The only exception to this rule is the M/G/infty.gif (58 bytes) queueing system. As a result of this negative side of the equivalence property, the presence in a queueing system of even a single facility with service times that are not negative exponential often creates serious analytical problems. 15

15 For practical purposes, however, the output of a M/G/m queueing system can often be assumed approximately Poisson for large m. This is true due to a limit theorem stating that, when a large number of renewal processes are pooled together, the resulting process approaches Poisson, irrespective of the type of the individual renewal processes.

14 The reader should be aware that many "subtleties" arise in the analysis not only of queueing networks that permit some types of feedback but also of some types of feedforward networks (see [KLEI 75] [KLEI 76] and, especially, [DISN 79]). For instance, in networks with feedback, the flow of units within the network may not be Poisson.