Statistical simulation of symmetric multiprocessor systems

Abstract

Statistical simulation is driven by a stream of randomly generated instructions, based on statistics collected during a single detailed simulation. This method can give accurate performance estimates within minutes, allowing a large design space to be simulated quickly. Prior work has applied this technique to superscalar processors. We evaluate the extension of statistical simulation to symmetric multiprocessing systems. Key program parameters are identified and program statistics are collected during detailed simulations for both multiprogrammed workloads (SpecInt) and parallel scientific workload (Splash-2). The accuracy of statistical simulation is evaluated at different levels of model detail, and it is shown that for multiprogrammed workloads a 10% average error can be achieved, and for parallel benchmark programs 15% average error can be achieved.