Predicting the performance of shared multiprocessor caches

Abstract

We investigate the performance of shared caches in a shared-memory multiprocessor executing parallel programs, and formulate simple models for estimating the load placed on the bus by such a shared cache. We analyze three parallel program traces to quantify the amount of sharing that takes place during program execution. These results indicate that shared caches can substantially reduce the load placed on a bus by a large number of processors. Finally we propose a new metric of group locality to aid in predicting the performance of a parallel program running on a machine with shared caches.