Opera: Similarity Analysis on Data Access Patterns of OpenMP Tasks to Optimize Task Affinity

OpenMP supports task-based parallelism, but task scheduling is oblivious to data locality, which leads to inconsistent performance. In this paper, we present Opera, an OpenMP task scheduler which leverages memory access information profiled offline to guide runtime task scheduling. The evaluation results show that Opera improves performance by up to 40% (21.2% on average), comparing with using three schedulers in the Nanos++ runtime library.