4 Customer's point of view

Let us now examine how these policies are perceived from the customer's side. Fig. 3 gives the cdf of the sojourn time as obtained from simulations, while Table 1 gives the dispersion parameters of the sojourn time for the eight policies considered in the preceding section.

Fig. 3: Cumulative distribution of the sojourn time

Table 1: Mean sojourn time and standard deviation

				mean	$\pm2\sigma$	ratio	sd	$\pm2\sigma$	ratio	$\tau_{0.1\%}$	$\tau/\mu$
	fast			470.53	7.00	1.00	445.61	10.01	1.00	3808	8.1
	load			586.72	7.36	1.25	463.88	10.60	1.04	4189	7.2
	jsiz			606.51	7.89	1.29	496.94	11.40	1.12	4435	7.4
	jran			675.53	7.09	1.44	670.32	11.94	1.50	7479	11.2
	size			621.48	7.69	1.32	500.77	10.48	1.12	4423	7.1
	rtwo			710.71	7.36	1.51	527.89	10.37	1.18	4817	6.8
	robn			997.21	13.15	2.12	904.20	18.61	2.03	8793	8.9
	rand			2102.66	31.03	4.47	1999.41	45.64	4.49	17560	8.4

A change of policy not only affects the mean sojourn time but also its dispersion : the columns labeled "ratio" are quite equal. Moreover, fat tails are appearing that worsen the behavior : in columns $\tau_{0.1\%}$ and $\tau/\mu$ are given the absolute and relative values of the last $1/1000$ fractile.

It must be emphasized that load policy is easy and costless to implement in a centralized environment : it suffices to pool the different queues into a single queue [9]. Additionally this organization provides a visible and provable individual fairness, especially when servers can be added or subtracted.