Producing Reliable Full-System Simulation Results: A Case Study of CMP with Very Large Caches

The greater detail and improved realism of fullsystem architecture simulation makes it a valuable computer architecture design tool. However, its unique characteristics introduce new sources of simulation variability which could make the results of such simulations less reliable. Meanwhile, the demand for more levels of cache and larger caches has increased to improve the system power and performance. This paper presents techniques to produce reliable results in fullsystem simulation of CMP computer systems with large caches. Specifically, we propose the detailed emulation replay warmup technique to deal with cold or incompletely warmed up large caches. We also propose the region of interest synchronization technique to prevent simulating non-representative phase when running multi-program workloads. Furthermore, we quantify the variation reduction one can achieve when using processor affinity and checkpointing. Finally, we show that by applying all four of these simulation techniques, the simulation variability is limited to less than 1% and the simulation results are therefore more reliable.