Simulation of multiprocessors: accuracy and performance

摘要: Simulation plays an important role in analyzing the performance of multiprocessor memory systems, but detailed simulation can consume enormous amounts CPU time. This dissertation investigates three techniques that reduce time required to simulate a system: trace-driven simulation, simplifying approximations memory-system simulators, and parallel simulation. Trace-driven allows multiple target systems from single interpretation workload, recorded trace. It is often assumed trace system-independent, this assumption violated by workloads have race-conditions or utilize dynamic scheduling. measures analyzes impact system-dependent traces on accuracy simulations multiprocessors. Two such for were investigated. Precise requires events originating different parts system be simulated correct order. The investigation revealed relaxing constraint double triple simulator with little loss accuracy. also simulating references private data, which effect performance. By neglecting references, was accelerated up 80 percent accuracy. To further increase speed, existing parallelized run DASH multiprocessor. challenge preserve without serializing simulation. Two approaches evaluated: (1) ordering synchronization operations only (2) order all small amount. Both yielded speedups, limited load imbalance. Overall, comparable hardware-assisted simulators as Wisconsin Wind Tunnel.