Hierarchical scheduling of DAG structured computations on manycore processors with dynamic thread grouping

  • Authors:

  • Yinglong Xia;Viktor K. Prasanna;James Li

  • Affiliations:

  • Department of Computer Science, University of Southern California, Los Angeles, CA;Department of Computer Science, University of Southern California, Los Angeles, CA and Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, CA;Ming Hsieh Department of Electrical Engineering, University of Southern California, Los Angeles, CA

  • Venue:
  • JSSPP'10 Proceedings of the 15th international conference on Job scheduling strategies for parallel processing
  • Year:
  • 2010

Quantified Score

Hi-index 0.01

Visualization

Abstract

Many computational solutions can be expressed as directed acyclic graphs (DAGs) with weighted nodes. In parallel computing, scheduling such DAGs onto manycore processors remains a fundamental challenge, since synchronization across dozens of threads and preserving precedence constraints can dramatically degrade the performance. In order to improve scheduling performance on manycore processors, we propose a hierarchical scheduling method with dynamic thread grouping, which schedules DAG structured computations at three different levels. At the top level, a supermanager separates threads into groups, each consisting of a manager thread and several worker threads. The supermanager dynamically merges and partitions the groups to adapt the scheduler to the input task dependency graphs. Through group merging and partitioning, the proposed scheduler can dynamically adjust to become a centralized scheduler, a distributed scheduler or somewhere in between, depending on the input graph. At the group level, managers collaboratively schedule tasks for their workers. At the within-group level, workers perform self-scheduling within their respective groups and execute tasks. We evaluate the proposed scheduler on the Sun UltraSPARC T2 (Niagara 2) platform that supports up to 64 hardware threads. With respect to various input task dependency graphs, the proposed scheduler exhibits superior performance when compared with other various baseline methods, including typical centralized and distributed schedulers.