Difference between revisions of "Sig Turbomachinery / Automatic parallel benchmarking"
From OpenFOAMWiki
m |
m |
||
Line 2: | Line 2: | ||
We have started using a dropletSplash test-case, developed by Hrvoje Jasak, and it has been run on several clusters. The major finding this far is that OpenFOAM scales linearly as long as the processes are distributed using a round-robin model, where multi-core nodes are not filled-up with processes from the same job. The interconnect is not a bottleneck as long as there are approximately 50,000 cells per process. | We have started using a dropletSplash test-case, developed by Hrvoje Jasak, and it has been run on several clusters. The major finding this far is that OpenFOAM scales linearly as long as the processes are distributed using a round-robin model, where multi-core nodes are not filled-up with processes from the same job. The interconnect is not a bottleneck as long as there are approximately 50,000 cells per process. | ||
+ | |||
+ | |||
+ | Back to [[Sig Turbomachinery / Work on specific projects|Work on specific projects]] |
Revision as of 18:23, 8 January 2008
A group is studying the scalability of OpenFOAM on different clusters, and for different configurations of OpenFOAM. The outcome of this work will be made public.
We have started using a dropletSplash test-case, developed by Hrvoje Jasak, and it has been run on several clusters. The major finding this far is that OpenFOAM scales linearly as long as the processes are distributed using a round-robin model, where multi-core nodes are not filled-up with processes from the same job. The interconnect is not a bottleneck as long as there are approximately 50,000 cells per process.
Back to Work on specific projects