Page 206 - A First Course In Stochastic Models
P. 206

SERVICE-SYSTEM DESIGN                     199

                Note that R is a dimensionless quantity that gives the average amount of work
                offered per time unit to the c servers. The offered load R is often expressed as R
                erlangs of work. In order to ensure the existence of a steady-state regime for the
                queue, it should be assumed that the service capacity c is larger than the offered
                load R. Hence the assumption is made that the server utilization
                                                  R
                                              ρ =
                                                   c
                is less than 1. Note that ρ represents the long-run fraction of time a given server is
                busy. In the single-server case the server utilization ρ should not be too close to 1
                in order to avoid excessive waiting of the customers. A rule of thumb for practical
                applications of the M/M/1 model is that the server utilization should not be much
                above 0.8. A natural question is how this rule of thumb should be adjusted for the
                multi-server case. It is instructive to have a look at Table 5.3.1. This table gives for
                several values of c and R the delay probability P W , the average waiting T W over
                the delayed customers and the 95% percentile η 0.95 of the steady-state waiting-time
                distribution of the delayed customers. In Table 5.3.1 we have normalized the mean
                service time 1/µ as 1. The delay probability P W (= P delay ) is given by formula
                (5.1.11). Since T W = W q /P W , it follows from (5.1.10) and (5.1.14) that

                                                   1
                                          T W =         .
                                               cµ(1 − ρ)
                By (5.1.10) and (5.1.13), the steady-state probability that a delayed customer has
                to wait longer than x time units is given by e −cµ(1−ρ)x  for x ≥ 0. Thus the pth
                percentile η p of the steady-state waiting-time distribution of the delayed customers
                is found from e −cµ(1−ρ)x  = 1 − p. This gives
                                        −1
                                η p =         ln(1 − p),  0 < p < 1.
                                     cµ(1 − ρ)
                  The following conclusion can be drawn from Table 5.3.1: high values of the
                server utilization ρ do not conflict with acceptable service to the customers when

                             Table 5.3.1  Service measures as function of c and R
                              ρ = R/c = 0.8   ρ = R/c = 0.95   ρ = R/c = 0.99
                             P W  T W  η 0.95  P W  T W  η 0.95  P W  T W  η 0.95
                     c = 1  0.8   5   14.98  0.95  20   59.91  0.99  100  299.6
                     c = 2  0.711  2.5  7.49  0.926 10  29.96  0.985  50  149.8
                     c = 5  0.554  1   3.0   0.878  4   11.98  0.975  20  59.91
                     c = 10  0.409  0.5  1.5  0.826  2  5.99  0.964  10  29.96
                     c = 25  0.209  0.2  0.6  0.728  0.8  2.40  0.942  4  11.98
                     c = 50  0.087  0.1  0.3  0.629  0.4  1.20  0.917  2  5.99
                     c = 100 0.020  0.05  0.15  0.506  0.2  0.60  0.883  1  3.0
                     c = 250 3.9E-4 0.02  0.06  0.318  0.08  0.24  0.818  0.4  1.2
                     c = 500 8.4E-7 0.01  0.03  0.177  0.04  0.12  0.749  0.2  0.6
   201   202   203   204   205   206   207   208   209   210   211