Reducing synchronization overhead in parallel simulation
PADS '96 Proceedings of the tenth workshop on Parallel and distributed simulation
Optimistic simulation of parallel message-passing applications
Proceedings of the fifteenth workshop on Parallel and distributed simulation
| Hi-index | 0.00 |
Synchronization is often the dominant cost in conservative parallel simulation, particularly in simulations of parallel computers, in which low-latency simulated communication requires frequent synchronization. This thesis presents local barriers and predictive barrier scheduling, two techniques for reducing synchronization overhead in the simulation of message-passing multicomputers. Local barriers use nearest-neighbor synchronization to reduce waiting time at synchronization points. Predictive barrier scheduling, a novel technique which schedules synchronizations using both compile-time and runtime analysis, reduces the frequency of synchronization operations. These techniques were evaluated by comparing their performance to that of periodic global synchronization. Experiments show that local barriers improve performance by up to 24% for communication-bound applications, while predictive barrier scheduling improves performance up to 65% for applications with long local computation phases. Because the two techniques are complementary, I advocate a combined approach. This work was done in the context of Parallel Proteus, a new parallel simulator of message-passing multicomputers.