Selecting a Single, Representative Sample for Accurate Simulation of SPECint Benchmarks

Abstract:

Detailed micro-architecture studies often require time-consuming,
cycle-accurate simulation. Unfortunately, many benchmarks of
interest, like the SPEC benchmarks, run for tens of billions of
instructions and are prohibitive to simulate to completion. Instead,
researchers typically must confine themselves to simulating samples of
the program's overall execution. This report presents detailed
interval-trace data about the SPECint programs' branch-prediction and
cache behavior. The report then argues that accurate simulation
results can be obtained using a single, short simulation window of
just 50 million instructions. They key to obtaining accurate results
is to choose this simulation window carefully, and in particular to
avoid unrepresentative behavior at the beginning of programs'
execution.