A Proof of Concept for Optimizing Task Parallelism by Locality Queues - Computer Science > PerformanceReportar como inadecuado




A Proof of Concept for Optimizing Task Parallelism by Locality Queues - Computer Science > Performance - Descarga este documento en PDF. Documentación en PDF para descargar gratis. Disponible también para leer online.

Abstract: Task parallelism as employed by the OpenMP task construct, although ideal fortackling irregular problems or typical producer-consumer schemes, bears somepotential for performance bottlenecks if locality of data access is important,which is typically the case for memory-bound code on ccNUMA systems. We presenta programming technique which ameliorates adverse effects of dynamic taskdistribution by sorting tasks into locality queues, each of which is preferablyprocessed by threads that belong to the same locality domain. Dynamicscheduling is fully preserved inside each domain, and is preferred overpossible load imbalance even if non-local access is required. The effectivenessof the approach is demonstrated using a blocked six-point stencil solver as atoy model.



Autor: Markus Wittmann, Georg Hager

Fuente: https://arxiv.org/







Documentos relacionados