Exploiting thread-data affinity in OpenMP with data access patterns

Abstract

In modern NUMA architectures, preserving data access locality is a key issue to guarantee performance. We define, for the OpenMP programming model, a type of architecture-agnostic programmer hint to describe the behaviour of parallel loops. These hints are only related to features of the program, in particular to the data accessed by each loop iteration. The runtime will then combine this information with architectural information gathered during its initialization, to guide task scheduling, in case of dynamic loop iteration scheduling. We prove the effectiveness of the proposed technique on the NAS parallel benchmark suite, achieving an average speedup of 1.21x. © 2011 Springer-Verlag.