Challenges of Reducing Cycle-Accurate Simulation Time for TBP Applications

Abstract

Cycle-accurate simulation is an important tool that depends on the computational power of supercomputers. Unfortunately, simulations of modern multi-core platforms can take weeks or months. In this paper, we look into the challenges of employing a sampling based technique for reducing simulation time of multi-threaded applications. We introduce FASTA, a simple 3- phase methodology for reducing the simulation time of Task Based Parallel (TBP) applications. FASTA takes advantage of the periodic behavior of parallel applications and identifies a small number of representative execution samples. By exploring a large design space we show that even though we can not use FASTA for every type of application, there are some for which a 12x speedup can be achieved with an accuracy error as low as 2.6%.