Performance Prediction of Task Workloads in Work-Stealing Runtimes for NUMA Multi-core Architectures

Abstract

Abstract: Work stealing is a popular load balancing technique which is successfully adapted by various user level task parallel runtime systems such as Cilk, TBB etc. On the other hand, processor manufacturers are working on releasing processors with more than one socket on chip, to overcome the memory wall problem. If work stealing based run-time systems are ported onto such multi-socket machines, it may lead to performance issues. This paper has attempted to study the impact of such multi-socket multi-core architectures on performance of work stealing based run-times. In this paper, a work stealing parameter called remote steal count is introduced specially for assessing work stealing run-times on multi-socket architectures. This paper also proposes a simple regression model between work stealing parameters and performance of task-based applications. This model is helpful to predict the performance impact of task parallel applications   while porting them into multi-socket hardware environment.