| Description |
Empirical performance evaluations of parallel, discrete event simulation algorithms using deadlock avoidance and deadlock detection and recovery techniques developed by Chandy and Misra have been performed using the BBN Butterfly multiprocessor. Experiments using synthetic workloads reveal that the degree to which processes can look ahead in simulated time plays a critical role in the performance of distributed simulators using these algorithms. These results are applied to a queueing network simulation where as much as an order of magnitude improvement in performance is observed if the distributed simulator is programmed to fully exploit the lookahead available in the application. Performance measurements of several hypercube-based communication network simulators provide additional empirical data to support these claims. These results demonstrate that substantial improvements in performance are obtainable if the application can be programmed to have good lookahead characteristics. On the other hand, other applications inherently contain poor lookahead properties, and appear to be ill-suited for these simulation algorithms. Keywords: Communications network simulation. Presented at the International Conference on Parallel Processing, Aug 1988. |
