Characterizing Rollbacks in Parallel Optimistic Simulations

Rollbacks are widely used to maintain causality parallel optimistic simulations, specifically in Time Warp synchronized simulations. Despite their significance, literature is scant on fundamental characterization of the two key metrics of rollbacks, namely - 1 inter-rollback cycles, i.e., how many event processing cycles elapse before a rollback occurs, and 2 rollback lengths, i.e., how many cycles does a rollback cancel. This study proposes an experimental method to characterize rollbacks via statistical analysis. We have conducted experimental analyses using a widely used synthetic benchmark called phold. We have conducted 1000s of simulations with different combinations of phold settings on two different computational-clusters to analyze rollback-profiles of a broad spectrum of parallel simulation configurations. Our analysis shows that both rollback metrics are geometrically distributed with their aggregate characteristics following a normal distribution. Interestingly, the overarching metrics from 500 different simulation configurations are also normally distributed.