Advanced Numerical Techniques of Performance Evaluation. Volume 2
WASHINGTON UNIV SEATTLE DEPT OF COMPUTER SCIENCE
Pagination or Media Count:
The conservative Bryant-Chandy-Misra parallel simulation paradigm was originally developed in a distributed system setting. However, the recent widespread availability of relatively inexpensive, medium-scale shared memory multiprocessors encourages a re-examination of traditional approaches to its implementation. Many of the obstacles to good performance, such as communication delay, sub-optimal model partitioning, artificial blocking, and the high cost of deadlock avoidance, deadlock detection, and deadlock recovery, can be reduced or even eliminated in a shared-memory implementation. This paper presents a number of techniques for efficient in a shared memory parallel simulation. All of these techniques have been implemented in Synapse, an integrated programming environmentrun-time system that supports efficient parallel discrete-event simulation on shared-memory multiprocessors. A new technique for artificial blocking and deadlock avoidance, lazy blocking avoidance, is built directly into Synapses multi-threaded run-time kernel. The Synapse run-time system also contains mechanisms for efficiently detecting and breaking the few deadlocks that are not prevented by lazy blocking avoidance. Our performance results are substantially better than those reported in previous empirical studies in this area, enough to challenge their negative conclusions about the feasibility of the conservative approach to parallel simulation.
- Computer Programming and Software