Self-similarity in SPLASH-2 workloads on shared memory multiprocessors systems

Abstract

The workload used for evaluating and obtaining performance results in shared memory multiprocessors are widely heterogeneous. Traces have been used over several decades and as computers systems grew in power, semantic benchmarks, like SPLASH2, became the most common workloads. Unfortunately, few benchmarks are available. Recently, self-similar studies have been performed in several computer domains. In this paper, we study the self-similar properties of several SPLASH2 benchmarks. Each benchmark has been studied independently, and all exhibit a clearly self-similar behaviour. The results enable the construction of a self-similar memory reference generator that makes a wide variety of parallel workload traces in a a flexible manner; as well as quickly.