Zero Lookahead? Zero Problem. The Window Racer Algorithm

Synchronization algorithms for parallel simulation struggle to attain speedup if the simulation entities are tightly coupled and their interactions are difficult to predict. Window Racer is a novel parallel synchronization algorithm for shared-memory architectures specifically targeted toward attaining speedup in these challenging cases. The key idea is to speculatively process sequences of dependent events even across partition boundaries through fine-grained locking and low-overhead rollbacks, while negotiating a global synchronization window that rules out transitive rollbacks. In performance measurements using a variant of the PHold benchmark model, Window Racer outperforms an established implementation of the Time Warp algorithm in model configurations where events are often scheduled with near-zero delay. In an ablation study, we pinpoint the performance impact of our algorithm’s individual features by reducing Window Racer to two existing algorithms. We further study the algorithm’s ability to attain speedup in simulations of bio-chemical reaction networks, a particularly challenging class of simulations with tightly coupled state transitions.