Distributed array data management on NUMA multiprocessors

Abstract
Management of program data to reduce false sharingand improve locality is critical for scaling performanceon NUMA multiprocessors. We use HPF-likedirectives to partition and place arrays in data-parallelapplications on Hector, a shared-memory NUMA multiprocessor.We present experimental results thatdemonstrate the magnitude of the performance improvementattainable when our proposed array managementschemes are used instead of the operating systemmanagement policies. We then describe a...

This publication has 3 references indexed in Scilit: