SimSnap: fast-forwarding via native execution and application-level checkpointing

Abstract

As systems become more complex, conducting cycle-accurate simulation experiments becomes more time consuming. Most approaches to accelerating simulations attempt to choose simulation points, such that the performance of the program portions modeled in detail are representative of whole-program behavior. To maintain or build the correct architectural state, "fast-forwarding" models a series of instructions before a desired simulation point. This fast-forwarding is usually performed by functional simulation: modeling the effects of instructions without all the details of pipeline stages and individual /spl mu/-ops. We present another fast-forwarding technique, SimSnap, that leverages native execution and application-level checkpointing. We demonstrate the viability of our approach by moving checkpointed versions of SPLASH-2 benchmarks between an Alpha 21264 system and SimpleScalar Version 4.0 Alpha-Sim. Reduction in experiment times is dramatic, with minimal perturbation of benchmark programs.