How Many Threads will be too Many? On the Scalability of OpenMP Implementations

Iwainsky, Christian; Knobloch, Michael; Calotoiu, Alexandru; Shudler, Sergei; Bischof, Christian; Wolf, Felix; Strube, Alexandre

doi:10.1007/978-3-662-48096-0_35

Contribution to a conference proceedings

FZJ-2016-00659

How Many Threads will be too Many? On the Scalability of OpenMP Implementations

Iwainsky, C. (Corresponding author) ; Shudler, S. ; Calotoiu, A. ; Strube, A.FZJ* ; Knobloch, M.FZJ* ; Bischof, C. ; Wolf, F.

2015
Springer Berlin Heidelberg Berlin, Heidelberg
ISBN: 978-3-662-48095-3 (print), 978-3-662-48096-0 (electronic)

Euro-Par 2015: Parallel Processing / Träff, Jesper Larsson (Editor)
21st International Conference on Parallel and Distributed Computing, Euro-Par 2015, Vienna, Austria, 24 Aug 2015 - 28 Aug 2015 Berlin, Heidelberg : Springer Berlin Heidelberg, Lecture Notes in Computer Science 9233, 451 - 463 (2015) [10.1007/978-3-662-48096-0_35]

This record in other databases:

Please use a persistent id in citations: doi:10.1007/978-3-662-48096-0_35

Abstract: Exascale systems will exhibit much higher degrees of parallelism both in terms of the number of nodes and the number of cores per node. OpenMP is a widely used standard for exploiting parallelism on the level of individual nodes. Although successfully used on today’s systems, it is unclear how well OpenMP implementations will scale to much higher numbers of threads. In this work, we apply automated performance modeling to examine the scalability of OpenMP constructs across different compilers and platforms. We ran tests on Intel Xeon multi-board, Intel Xeon Phi, and Blue Gene with compilers from GNU, IBM, Intel, and PGI. The resulting models reveal a number of scalability issues in implementations of OpenMP constructs and show unexpected differences between compilers.

Classification: