ALWP: A Workload Partition Method for the Efficient Parallel Simulation of Manycores

this paper addresses the workload partition strategies in simulating many-core architectures. The key observation behind this paper is: compared to multicore, manycore features with more non-uniform memory access and unpredictable network traffic; these features degrade simulation speed and accuracy of parallel discrete event simulators (PDES) in cases of static workload partition schemes. Based on the observation, an adaptive workload partition method is proposed: Adaptive Link-based Workload Partition (ALWP). The proposed method can deliver more speedup and accuracy than traditional static partition schemes. It achieves this ability by placing the partitioning border prior on less interactive links. Using a PDES simulator, we evaluate the performance of ALWP in simulating a 256-core general purpose many-core processor. Running SPLASH2 benchmark applications, the experimental results demonstrate that, compared to static scheme, ALWP can deliver speed improvement by 14%~42% and reduce more timing errors (>50%) in relaxed cases.