Sunday, August 12, 2012

Connecting Two Posts: AMI Parallelizability and Hyperthreading Investiagtion

The most recent post shows that parallel efficiency stays constant with any more than one core with AMI. In an older post testing the effect of hyperthreading, efficiency went down, instead of staying constant. These results are still consistent, because the two tests were fundamentally different, one obviously using hyperthreading and the other using true parallelism. Hyperthreading makes more virtual cores, but with less physical resources per core (makes sense; you cannot get something for nothing). Since we hypothesized that the lower efficiency is due to serial component, having weaker cores will lead to a slower overall simulation, since the serial component (which can only be run on one processor) will have less resources. Thus the lower performance using hyperthreading is due to the serial component running on a weaker single core. Also, the software could have made a difference as that test was run with GGI on OpenFOAM 1.5-dev.

This has practical implications for building / choosing a computer to use with AMI simulations. One would want to choose a processor with the strongest single-core performance. Choosing a processor with larger number of cores, but weaker single cores would have a much more detrimental effect in AMI parallel simulations than one would think without considering this post. This is actually applicable to all simulations as there is always some serial component, but the lower the parallel efficiency, the more important the aforementioned guideline is.

No comments:

Post a Comment