摘要：Software architecture simulators are indispensable tools in modern processor design. According to the granularity of simulation, they can be classified into the fast functional simulation and the slow detailed one. The detailed simulator takes far longer time than the functional simulator when simulating the same workload. Based on the duration difference of them, we propose a Workload Segmented Parallel Simulation (WSPS) methodology to accelerate the detailed simulation by simulating different segments of the workload concurrently. The results on SPEC2Kint benchmarks show that, when programs are divided into 64 segments, the speedup is about 11.5, with the relative error of CPI and L1 cache hit-rate remaining lower than 1.5% and 0.01%, respectively. Also, the analysis indicates that WSPS-based simulation can achieve even much higher speedup when using more complicated simulation models, and its duration can approach that of the functional simulation with the accuracy remaining acceptable if the workload size is large enough.
