Overcoming over-optimism in Time Warp via aggregation of fast processes
Sachdev, Vinay Suresh
MetadataShow full item record
The parallel Time Warp protocol is widely used to improve the performance of large-scale discrete event simulations such as simulations of the air traffic control system and the World Wide Web. However, a concern regarding optimistic simulation processing is that some logical processes (LPs) may progress far beyond others into the simulated future, causing an imbalance that may degrade performance. The degradation is due to long and excessive rollbacks, inefficient use of memory resources, and communication overheads. Allowing this “over-optimistic” processing may bring the simulation to a halt. This is unacceptable for long running simulations that may take days to complete. We present a new mechanism that controls over-optimistic processing through aggregation and isolation of fast processes and redistribution of the slow, less optimistic processes. Our performance results demonstrate that our techniques can improve the useful work by a factor of 1.75 and also improve execution time, while keeping the overhead small.