Cooperative Write-Behind Data Buffering for MPI I/O

  • Wei-keng Liao
  • Kenin Coloma
  • Alok Choudhary
  • Lee Ward
Conference paper

DOI: 10.1007/11557265_17

Part of the Lecture Notes in Computer Science book series (LNCS, volume 3666)
Cite this paper as:
Liao W., Coloma K., Choudhary A., Ward L. (2005) Cooperative Write-Behind Data Buffering for MPI I/O. In: Di Martino B., Kranzlmüller D., Dongarra J. (eds) Recent Advances in Parallel Virtual Machine and Message Passing Interface. EuroPVM/MPI 2005. Lecture Notes in Computer Science, vol 3666. Springer, Berlin, Heidelberg

Abstract

Many large-scale production parallel programs often run for a very long time and require data checkpoint periodically to save the state of the computation for program restart and/or tracing the progress. Such a write-only pattern has become a dominant part of an application’s I/O workload and implies the importance of its optimization. Existing approaches for write-behind data buffering at both file system and MPI I/O levels have been proposed, but challenges still exist for efficient design to maintain data consistency among distributed buffers. To address this problem, we propose a buffering scheme that coordinates the compute processes to achieve the consistency control. Different from other earlier work, our design can be applied to files opened in read-write mode and handle the patterns with mixed MPI collective and independent I/O calls. Performance evaluation using BTIO and FLASH IO benchmarks is presented, which shows a significant improvement over the method without buffering.

Keywords

Write behind MPI I/O file consistency data buffering I/O thread 

Preview

Unable to display preview. Download preview PDF.

Unable to display preview. Download preview PDF.

Copyright information

© Springer-Verlag Berlin Heidelberg 2005

Authors and Affiliations

  • Wei-keng Liao
    • 1
  • Kenin Coloma
    • 1
  • Alok Choudhary
    • 1
  • Lee Ward
    • 2
  1. 1.Electrical and Computer Engineering DepartmentNorthwestern University 
  2. 2.Scalable Computing Systems DepartmentSandia National Laboratories 

Personalised recommendations