Hybrid MPI-cell parallelism for hyperbolic PDE simulation on a cell processor cluster

Scott Rostrup, Hans De Sterck

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

Abstract

We show how a numerical simulation method for nonlinear hyperbolic partial differential equation (PDE) systems on structured grids with explicit timestepping can be implemented efficiently for the Cell processor and for clusters of Cell processors. We describe memory layout, communication patterns and optimization steps that are performed to exploit the parallel architecture of the Cell processor. A second layer of Message Passing Interface (MPI) parallelism is added to obtain a hybrid parallel code that can be executed efficiently on Cell clusters. Performance tests are conducted on a Cell cluster, and the Cell performance is compared with x86 performance (Xeon). Compared with single-core Xeon performance, the Cell processor obtains significant speed-ups of 60x for single precision calculations, and 20x for double precision. In a chip-to-chip comparison, the Cell code is 14x faster than a 4-core Xeon (using pthreads) in single precision, and 5x faster in double precision. Parallel cluster scaling results were hampered by a relatively slow interconnect on our test system, but overall our study shows how Cell clusters can be used efficiently for simulating nonlinear hyperbolic PDE systems.

Original languageEnglish
Title of host publicationHigh Performance Computing Systems and Applications - 23rd International Symposium, HPCS 2009, Revised Selected Papers
Pages337-348
Number of pages12
Volume5976 LNCS
DOIs
Publication statusPublished - 2010
Externally publishedYes
Event23rd International Symposium on High Performance Computing Systems and Applications - Kingston, ON, Canada
Duration: 14 Jun 200917 Jun 2009
Conference number: 23

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume5976 LNCS
ISSN (Print)03029743
ISSN (Electronic)16113349

Conference

Conference23rd International Symposium on High Performance Computing Systems and Applications
CountryCanada
CityKingston, ON
Period14/06/0917/06/09

Cite this