Understanding the causes of performance variability in HPC workloads

David Skinner, William Kramer

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

While most workload characterization focuses on application and architecture performance, the variability in performance also has wide ranging impacts on the users and managers of large scale computing resources. Performance variability, though secondary to absolute performance itself, can significantly detract from both the overall performance realized by parallel workloads and the suitability of a given architecture for a workload. In making choices about how to best match an HPC workload to an HPC architecture most examinations focus primarily on application performance, often in terms nominal or optimal performance. A practical concern which brackets the degree to which one can expect to see this performance in a multi-user production computing environment is the degree to which performance varies. Without an understanding of the performance variability exhibited by a computer for a given workload, in a practical sense, the effective performance that can be realized is still undetermined. In this work we examine both architectural and application causes of variability, quantify their impacts, and demonstrate performance gains realized by reducing variability.

Original languageEnglish (US)
Title of host publicationProceedings of the 2005 IEEE International Symposium on Workload Characterization, IISWC-2005
Pages137-149
Number of pages13
DOIs
StatePublished - Dec 1 2005
Externally publishedYes
Event2005 IEEE International Symposium on Workload Characterization, IISWC-2005 - Austin, TX, United States
Duration: Oct 6 2005Oct 8 2005

Publication series

NameProceedings of the 2005 IEEE International Symposium on Workload Characterization, IISWC-2005
Volume2005

Other

Other2005 IEEE International Symposium on Workload Characterization, IISWC-2005
CountryUnited States
CityAustin, TX
Period10/6/0510/8/05

ASJC Scopus subject areas

  • Engineering(all)

Fingerprint Dive into the research topics of 'Understanding the causes of performance variability in HPC workloads'. Together they form a unique fingerprint.

  • Cite this

    Skinner, D., & Kramer, W. (2005). Understanding the causes of performance variability in HPC workloads. In Proceedings of the 2005 IEEE International Symposium on Workload Characterization, IISWC-2005 (pp. 137-149). [1526010] (Proceedings of the 2005 IEEE International Symposium on Workload Characterization, IISWC-2005; Vol. 2005). https://doi.org/10.1109/IISWC.2005.1526010