The authors present an analytical model and corroborating experimental results that illustrate the relationship between two windowing protocols for the synchronization of parallel and distributed discrete event simulations. Windowing protocols are a compromise between traditional optimistic and conservative synchronization mechanisms. Essentially, windowing protocols throttle optimistic execution by incorporating barrier synchronizations and global reductions on functions of future simulation times.
Both bounded time warp (BTW) and YAWNS define the execution of a logical process (LP) in three phases. A global window of simulation time is cooperatively defined in the first phase. In the second phase, LPs concurrently execute events whose timestamps fall within the window. The third phase consists of a barrier synchronization, and then the process repeats. The primary difference between the two protocols is the window size. YAWNS windows are small enough to guarantee that all processing within the same window will be correct. BTW defines larger windows, and allows LPs to optimistically process all events within the window.
The model predicts, and the experiments verify, an optimally sized window for BTW. The window is much larger than that for YAWNS, but LPs generally execute only one or two events within this larger window. When the number of LPs per physical processor is low, BTW outperforms YAWNS. As the number of LPs per physical processor increases, YAWNS can outperform BTW.
The analytic model developed in the paper represents an excellent, rigorous, and occasionally novel approach. The results add to the growing body of evidence that the performance of parallel simulations is less a function of the specific synchronization protocols than of state saving and communication costs, problem size, and load distribution.