An Empirical Evaluation of Virtual Circuit Holding Times in IP-over-ATM Networks

of network pricing. We collected data to empirically model the workload. The data shows temporal locality of reference in all traces. LRU based holding time policies use this temporal locality eeectively and perform better than any other strategy examined. To our surprise, timer based schemes perform poorly. However, the LRU-delay scheme, which corresponds to a xed system-wide timeout, is eeective and trivial to implement. We have studied a number of variations of holding time policies and have found that none of the variations ooer signiicant improvements in performance over the basic strategies discussed here. 7 Acknowledgements We wish to thank Bruce Mah at UC Berkeley and Peter Danzig at the University of Southern Califor-nia for helping collect the traces. Thanks also to Ra-mon Caceres at Matsushita Labs, who suggested the relevance of pricing in studying holding times, and made several useful comments on an earlier draft of this work. We varied from 0:25 to 8:0 and computed the total cost incurred for each dataset using each of the holding policies (here, for the LRU-delay based policy, we use c = 2). In Figure 5 we plot the variation in cost as is varied for a sample dataset (the other datasets showed similar behavior). The cost increases as is increased. This is to be expected as a larger implies that each open costs much more. Even though the number of opens decrease when is increased, the holding time costs go up while the costs associated with opens do not decrease. The non-causal strategy is signiicantly better than any of the others though the gap narrows as decreases. To compare relative performance of the strategies across diierent values of and diierent datasets, we normalize the cost by dividing by the cost of the non-causal optimal strategy on the same data. A sample of the resulting values are shown in Figure 6. Table 2 lists the average taken over the all traces of the normalized cost of each policy. Timeout is consistently the worst, though its relative performance is not as bad as in the paging model. The LRU-delay based policy is the best, and LRU-statistics performs reasonably well. We see that the performance of LRU based policies improves as is increased whereas that of Timeout rst degrades and then improves. The LRU-statistics scheme performs quite poorly for values of < 1 while in the range 1 4 the performance is within a …