Updates on Tests on Intel® Multicore Processors
With every release of COMSOL Multiphysics, we run a series of tests on Intel® multicore processors to benchmark our utilization of the parallel processing capabilities.
Screenshot from the Intel® Concurrency Checker having just run a test of the stress analysis of a car wheel in COMSOL Multiphysics version 4.3. Enlarge image.
Intel® helpfully provides a tool that allows companies like ours to measure this utilization. Known as the Intel® Concurrency Checker, it measures how well the threads are being processed and utilized, by COMSOL Multiphysics, and returns a value for Computed Scaling. This is the number of threads that was possible to be utilized during the test. A score of 1.0 means that basically no parallelization occurred as more than one core was not accessed or utilized during the computational processing. On the other end of this scale, it is impossible to score more than 8.0 with 8 concurrent multicores. It should be pointed out though that these scores only relate to how many of the cores are utilized at the same time, and cannot be directly construed to indicate improvements in speedup.
Last time I reported our results, we achieved a score of 4.41. This was in excess of their recommendations and put us almost in the top decile of those that were accessing the Concurrency Checker.
This time around, we pushed this value up to 7.15. A great improvement! As Intel® recommends a minimum value of 2.6 for Computed Scaling, this score puts us way ahead of their recommendations. It also gives an indication to the underlying improvements, at least for parallel processing, to the core coding that has occurred between COMSOL Multiphysics version 4.2a and version 4.3.
September 17, 2012 at 3:38 am
Hi Phil! Quite useful info you shared. However I wanted some information regarding COMSOL in terms of memory utilization. We have some models which require more memory then we have and we can not just keep increasing it. I have been wondering if COMSOL supports model partitioning at the start and solve them one after another individually and iteratively. This would require much larger time for running the model though but we can afford it. please let me know if there is any possibility of this kind in COMSOL 4.3. You may direct me to one of your colleagues if this is not your domain. Thanks.
September 18, 2012 at 3:04 pm
I don’t know about the model partitioning – I don’t recognise the term. I suggest you send that question to our Support Team; email@example.com. But I do know that we support distribution on clusters. I’m not sure whether this would help you
Login to Comment
- Create Account
- Forgot your password?