Communication-aware Parallel Domain Decomposition using Space Filling Curves

Sudip Seal, Srinivas Aluru

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

1 Scopus citations

Abstract

Space filling curves (SFCs) are proximity preserving linearizations of multidimensional data. They are frequently used for parallel domain decomposition in scientific computing applications where interactions occur between elements with physical proximity. The conventional way to affect SFC-based parallel domain decomposition is to take the SFC-linearization and map it to processors in a block distributed fashion. In this paper, we derive closed form formulas for the average nearest neighbor distance along a Z-space filling curve and use this result to show that such a mapping may not be communication efficient. We introduce topology-aware mapping of SFC data to parallel computers and show that SFC data can be partitioned on hypercubes such that communication is restricted to nearest neighbors.

Original languageEnglish
Title of host publication19th International Conference on Parallel and Distributed Computing Systems 2006, PDCS 2006
PublisherInternational Society for Computers and Their Applications (ISCA)
Pages159-164
Number of pages6
ISBN (Electronic)9781604236446
StatePublished - 2006
Externally publishedYes
Event19th International Conference on Parallel and Distributed Computing Systems, PDCS 2006 - San Francisco, United States
Duration: Sep 20 2006Sep 22 2006

Publication series

Name19th International Conference on Parallel and Distributed Computing Systems 2006, PDCS 2006

Conference

Conference19th International Conference on Parallel and Distributed Computing Systems, PDCS 2006
Country/TerritoryUnited States
CitySan Francisco
Period09/20/0609/22/06

Keywords

  • domain decomposition
  • parallel algorithms
  • parallel communication
  • space filling curves
  • topology aware mapping

Fingerprint

Dive into the research topics of 'Communication-aware Parallel Domain Decomposition using Space Filling Curves'. Together they form a unique fingerprint.

Cite this