Accurately measuring MPI broadcasts in a computational grid

PDF Version Also Available for Download.

Description

An MPI library's implementation of broadcast communication can significantly affect the performance of applications built with that library. In order to choose between similar implementations or to evaluate available libraries, accurate measurements of broadcast performance are required. As we demonstrate, existing methods for measuring broadcast performance are either inaccurate or inadequate. Fortunately, we have designed an accurate method for measuring broadcast performance, even in a challenging grid environment. Measuring broadcast performance is not easy. Simply sending one broadcast after another allows them to proceed through the network concurrently, thus resulting in inaccurate per broadcast timings. Existing methods either fail to ... continued below

Physical Description

96 Kilobytes pages

Creation Information

T, Karonis N & de Supinski, B R May 6, 1999.

Context

This article is part of the collection entitled: Office of Scientific & Technical Information Technical Reports and was provided by UNT Libraries Government Documents Department to Digital Library, a digital repository hosted by the UNT Libraries. More information about this article can be viewed below.

Who

People and organizations associated with either the creation of this article or its content.

Sponsor

Publisher

Provided By

UNT Libraries Government Documents Department

Serving as both a federal and a state depository library, the UNT Libraries Government Documents Department maintains millions of items in a variety of formats. The department is a member of the FDLP Content Partnerships Program and an Affiliated Archive of the National Archives.

Contact Us

What

Descriptive information to help identify this article. Follow the links below to find similar items on the Digital Library.

Description

An MPI library's implementation of broadcast communication can significantly affect the performance of applications built with that library. In order to choose between similar implementations or to evaluate available libraries, accurate measurements of broadcast performance are required. As we demonstrate, existing methods for measuring broadcast performance are either inaccurate or inadequate. Fortunately, we have designed an accurate method for measuring broadcast performance, even in a challenging grid environment. Measuring broadcast performance is not easy. Simply sending one broadcast after another allows them to proceed through the network concurrently, thus resulting in inaccurate per broadcast timings. Existing methods either fail to eliminate this pipelining effect or eliminate it by introducing overheads that are as difficult to measure as the performance of the broadcast itself. This problem becomes even more challenging in grid environments. Latencies a long different links can vary significantly. Thus, an algorithm's performance is difficult to predict from it's communication pattern. Even when accurate pre-diction is possible, the pattern is often unknown. Our method introduces a measurable overhead to eliminate the pipelining effect, regardless of variations in link latencies. choose between different available implementations. Also, accurate and complete measurements could guide use of a given implementation to improve application performance. These choices will become even more important as grid-enabled MPI libraries [6, 7] become more common since bad choices are likely to cost significantly more in grid environments. In short, the distributed processing community needs accurate, succinct and complete measurements of collective communications performance. Since successive collective communications can often proceed concurrently, accurately measuring them is difficult. Some benchmarks use knowledge of the communication algorithm to predict the timing of events and, thus, eliminate concurrency between the collective communications that they measure. However, accurate event timing predictions are often impossible since network delays and local processing overheads are stochastic. Further, reasonable predictions are not possible if source code of the implementation is unavailable to the benchmark. We focus on measuring the performance of broadcast communication.

Physical Description

96 Kilobytes pages

Source

  • Eighth International Symposium on High-Performance Distributed Computing, Redondo Beach, CA (US), 08/03/1999--08/06/1999

Language

Item Type

Identifier

Unique identifying numbers for this article in the Digital Library or other systems.

  • Report No.: UCRL-JC-133177--Rev-1
  • Report No.: DP0101031
  • Grant Number: W-7405-ENG-48
  • Office of Scientific & Technical Information Report Number: 12133
  • Archival Resource Key: ark:/67531/metadc625016

Collections

This article is part of the following collection of related materials.

Office of Scientific & Technical Information Technical Reports

What responsibilities do I have when using this article?

When

Dates and time periods associated with this article.

Creation Date

  • May 6, 1999

Added to The UNT Digital Library

  • June 16, 2015, 7:43 a.m.

Description Last Updated

  • May 5, 2016, 9 p.m.

Usage Statistics

When was this article last used?

Yesterday: 0
Past 30 days: 1
Total Uses: 4

Interact With This Article

Here are some suggestions for what to do next.

Start Reading

PDF Version Also Available for Download.

Citations, Rights, Re-Use

T, Karonis N & de Supinski, B R. Accurately measuring MPI broadcasts in a computational grid, article, May 6, 1999; California. (digital.library.unt.edu/ark:/67531/metadc625016/: accessed September 25, 2017), University of North Texas Libraries, Digital Library, digital.library.unt.edu; crediting UNT Libraries Government Documents Department.