Limits of Work-Stealing Scheduling

Abstract

Abstract. The number of applications with many parallel cooperating processes is steadily increasing, and developing efficient runtimes for their execution is an important task. Several frameworks have been developed, such as MapReduce and Dryad, but developing scheduling mechanisms that take into account processing and communication requirements is hard. In this paper, we explore the limits of work stealing scheduler, which has empirically been shown to perform well, and evaluate loadbalancing based on graph partitioning as an orthogonal approach. All the algorithms are implemented in our Nornir runtime system, and our experiments on a multi-core workstation machine show that the main cause of performance degradation of work stealing is when very little processing time, which we quantify exactly, is performed per message. This is the type of workload in which graph partitioning has the potential to achieve better performance than work-stealing.

A Simple Improvement of the Work-stealing Scheduling Algorithm

“…We have found that the workstealing scheduler yields superior performance as long as the ratio of useful work and context-switch overhead is ≥ 75 [7].…”

Hypergraph Partitioning through Vertex Separators on Graphs

“…The real impact of these works turned out to be the introduction of hypergraph models to the combinatorial scientific computing community. Since then, the modeling power of hypergraphs appealed to many researchers and was applied to a wide variety of parallel and distributed computing applications such as data aggregation [15], image-space parallel direct volume rendering [7], parallel mixed integer linear programming [53], data declustering for multi-disk databases [38,42], scheduling file-sharing tasks in heterogeneous masterslave computing environments [33,34,37], and work-stealing scheduling [60]. Hy-pergraphs were also applied to applications outside the parallel computing domain such as road network clustering for efficient query processing [19,20,21], patternbased data clustering [43], reducing software development and maintenance costs [4], processing spatial join operations [51], and improving locality in memory or cache performance [1,52,61].…”