Finding probably better system configurations quickly

Abstract

The performance of computer and communication systems can in theory be optimized by iteratively finding better system configurations. However, a bottleneck is the time required in simulations/experiments for finding a better system configuration in each iteration. We propose algorithms that quickly find a system configuration that is probably better than the "standard" system configuration, where the performance of a given system configuration is estimated via simulations or experiments. We prove that our algorithms make correct decisions with high probability, and various heuristics to reduce the total simulation time are proposed. Numerical experiments show the effectiveness of the proposed algorithms, and this leads to several guidelines for designing efficient and reliable optimization procedures for the performance of computer and communication systems. Copyright 2006 ACM.