Going Native: Faster Architectural Simulation Fast-Forwarding

As system complexity grows, cycle-accurate simulation experiments become inordinately time consuming. Most approaches to accelerating architectural simulation model only portions of an application in detail. Detailed simulation begins with architectural state from earlier portions of the execution. Constructing this requires fast-forwarding: modeling effects of earlier instructions without modeling full architectural details. Fast-forwarding usually employs functional simulation within the cycle-accurate model. We present a rapid fast-forwarding technique, NativeFF, that leverages native execution and Application-Level Checkpointing. We demonstrate its viability by moving checkpointed state for a range of applications between native execution on an Alpha 21264 system and simulated execution in SimpleScalar Version 4.0. Our approach is orthogonal to—and combines well with—other approaches for reducing design space explorations, including SimPoint, SMARTS, reduced input sets, and Plackett-Burman or other statistical evaluations to reduce the experiment space required to generate sufficiently reliable results. Our experiments demonstrate that NativeFF reduces experiment time dramatically with minimal application perturbation.