Configurable Strategies for Work-stealing

TL;DR

This paper introduces configurable scheduling strategies for work-stealing systems, enabling applications to provide task-specific hints that optimize execution order and improve performance across various benchmarks.

Contribution

It proposes a novel approach allowing applications to dynamically influence task scheduling, enhancing flexibility, composability, and efficiency in work-stealing systems.

Findings

Prioritization reduces total work in branch-and-bound and shortest path algorithms.

Dynamic merging and half-work stealing improve performance.

Combining strategies enhances scheduling flexibility and efficiency.

Abstract

Work-stealing systems are typically oblivious to the nature of the tasks they are scheduling. For instance, they do not know or take into account how long a task will take to execute or how many subtasks it will spawn. Moreover, the actual task execution order is typically determined by the underlying task storage data structure, and cannot be changed. There are thus possibilities for optimizing task parallel executions by providing information on specific tasks and their preferred execution order to the scheduling system. We introduce scheduling strategies to enable applications to dynamically provide hints to the task-scheduling system on the nature of specific tasks. Scheduling strategies can be used to independently control both local task execution order as well as steal order. In contrast to conventional scheduling policies that are normally global in scope, strategies allow the scheduler to apply optimizations on individual tasks. This flexibility greatly improves composability as it allows the scheduler to apply different, specific scheduling choices for different parts of applications simultaneously. We present a number of benchmarks that highlight diverse, beneficial effects that can be achieved with scheduling strategies. Some benchmarks (branch-and-bound, single-source shortest path) show that prioritization of tasks can reduce the total amount of work compared to standard work-stealing execution order. For other benchmarks (triangle strip generation) qualitatively better results can be achieved in shorter time. Other optimizations, such as dynamic merging of tasks or stealing of half the work, instead of half the tasks, are also shown to improve performance. Composability is demonstrated by examples that combine different strategies, both within the same kernel (prefix sum) as well as when scheduling multiple kernels (prefix sum and unbalanced tree search).