Integrated Approach to Improving Performance of MPI Programs

The concept of “Beowulf” clusters, or cheap clusters formed by a network of workstations, increased dramatically the performance-over-price ratio in the high performance computing world. Nevertheless, many domain scientists and engineers using such clusters have realized that the performance they can actually harness is only a fraction of what is available. Two reasons are mainly responsible for this effect. First, the level of complexity required for the communication of an application to be optimized for a particular cluster makes such optimizations out of reach for most non-expert systems programmers. At the same time the lack of portability experienced by applications tuned for a specific cluster led most domain scientists and engineers to follow high-level generic communication solutions in their applications. In this project we use compiler technology to automatically fine-tune applications while preserving the simplicity experienced by the programmer.

Some of the research contributions have been:

  • Identify and evaluate transformations that can be applied to scientific applications in order to minimize communication overheads through communication-computation overlap.
  • Identify the performance metrics that characterize a cluster and develop a set of tools to actively probe them.
  • Develop a source-to-source optimizing system that can utilize the known information about a cluster and automatically apply the optimal transformations on scientific applications.
  • Develop a communication library that works in concert with the MPI library to increase communication-computation overlap in MPI programs.

Subprojects

ASPHaLT - Anthony Danalis, Lori Pollock, Martin Swany, and John Cavazos. “MPI-aware compiler optimizations for improving communication-computation overlap.” International Conference on Supercomputing (ICS'09), June 2009.

Gravel - Anthony Danalis, Aaron Brown, Lori Pollock, Martin Swany, and John Cavazos. “Gravel: A Communication Library to Fast Path MPI.” EuroPVM/MPI 2008, LNCS 5205, Springer-Verlag Berlin Heidelberg, pp. 111–119. September 2008.

Publications

Contributors

  • Faculty: Lori Pollock, Martin Swany
  • Ph.D. Students: Antonios Danalis
  • Collaborators: John Cavazos, University of Delaware
  • Past Undergraduate Students: Lewis Fishgold, Ki-Yong Kim, Aaron Brown
projects/asphalt.txt · Last modified: 2012/01/23 17:24 by sprenkle
  • 213 Smith Hall   •   Computer & Information Sciences   •   Newark, DE 19716  •   USA
    Phone: 302-831-6339  •   Fax: 302-831-8458