Performance Modeling of Stencil Computation on SW26010 Processors

Yao Liu, Li Liu, Mengtao Hu, Wei Wang, Wei Xue, Qingting Zhu

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

1 Scopus citations

Abstract

Stencil computation is a basic part in a large variety of scientific computing programs, especially for those containing partial differential equations. Due to the limited memory bandwidth, it is a challenge to improve the parallel efficiency of stencil computation on modern supercomputers. Performance modeling is the most common method of performance analysis. In this paper, we propose the generic performance model based on Sunway TaihuLight which is powered by SW26010 heterogeneous many-core processors. The generic model indicates the interaction between the programs and the computing platform from the architecture perspective, and points out the performance bottlenecks of the programs from the optimization perspective. Furthermore, we propose the specific performance model of stencil computation on SW26010 processors, and optimize the performance of stencil computation under the guidance of the model. The experimental results show that the performance models proposed in this paper are effective—the average error ratio of the predicted performance is less than 7%. Guided by the specific model, the optimized stencil computation achieves better performance than the unoptimized many-core version by 154.71% on 4096 cores.

Original languageEnglish
Title of host publicationAlgorithms and Architectures for Parallel Processing - 20th International Conference, ICA3PP 2020, Proceedings
EditorsMeikang Qiu
PublisherSpringer Science and Business Media Deutschland GmbH
Pages386-400
Number of pages15
ISBN (Print)9783030602444
DOIs
StatePublished - 2020
Event20th International Conference on Algorithms and Architectures for Parallel Processing, ICA3PP 2020 - New York, United States
Duration: 2 Oct 20204 Oct 2020

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume12452 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference20th International Conference on Algorithms and Architectures for Parallel Processing, ICA3PP 2020
Country/TerritoryUnited States
CityNew York
Period2/10/204/10/20

Keywords

  • Heterogeneous many-core processors
  • Performance modeling
  • Stencil computation
  • Sunway TaihuLight

Fingerprint

Dive into the research topics of 'Performance Modeling of Stencil Computation on SW26010 Processors'. Together they form a unique fingerprint.

Cite this