Performance measurements of distributed simulation strategies

A multiprocessor-based, distributed simulation testbed is described that facilitates controlled experimentation with distributed simulation algorithms. The performance of simulation strategies using deadlock avoidance and deadlock detection and recovery techniques are examined using various synthetic and actual workloads. The distributed simulators are compared with a uniprocessor-based event list implementation. Results of a series of experiments demonstrate that message population and the degree to which processes can look ahead in simulated time play critical roles in the performance of distributed simulators using these algorithms. An "avalanche" phenomenon was observed in the deadlock detection and recovery simulator, and was found to be a necessary condition for achieving good performance. The central server queueing model was also examined. The poor behavior of this test case that has been observed by others is reproduced in the testbed, and explained in terms of message population and lookahead. Based on these observations, a modification to the server process program is suggested that improves performance by as much as an order of magnitude when firstcome- first-serve (FCFS) servers are used. These results demonstrate that conservative distributed simulation algorithms using deadlock avoidance or detection and recovery techniques can provide significant speedups over sequential event list implementations for some workloads, even in the presence of only a moderate amount of parallelism and many feedback loops. However, a moderate to high degree of parallelism is not sufficient to guarantee good performance.