Menu

Discussion Forum

New Discussion
Filter

Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.

Increasing the Speed of A Time-dependent Study by Using Cluster Computing

Posted Jan 23, 2020, 6:48 a.m. EST Cluster & Cloud Computing Version 5.4 2 Replies

Eman Salman
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Hi,

I've seen some posts about using the cluster to speed-up the studies with parametric sweeps, but how can I use the cluster effectively to speed-up a time-dependent study?

Thanks,

Eman

2 Replies Last Post Jan 23, 2020, 5:06 p.m. EST
Edgar J. Kaiser Certified Consultant
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 4 years ago Jan 23, 2020, 8:27 a.m. EST

Eman,

this is an interesting question I had been struggling with as well. While parametric sweeps may run independently for each parameter value and can thus be efficiently distributed on cluster nodes this is not the case for the time steps in a transient study. So as long as your model fits into one node memory wise, the cluster will probably add overhead. Now, I have to admit that I don't have a cluster available and so I can't test and challenge this hypothesis. Would be interesting to hear other's experiences and opinions.

Cheers Edgar

-------------------
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
Eman, this is an interesting question I had been struggling with as well. While parametric sweeps may run independently for each parameter value and can thus be efficiently distributed on cluster nodes this is not the case for the time steps in a transient study. So as long as your model fits into one node memory wise, the cluster will probably add overhead. Now, I have to admit that I don't have a cluster available and so I can't test and challenge this hypothesis. Would be interesting to hear other's experiences and opinions. Cheers Edgar
Jim Freels mechanical side of nuclear engineering, multiphysics analysis, COMSOL specialist
Send Private Message Flag post as spam

Please login with a confirmed email address before reporting spam

Posted: 4 years ago Jan 23, 2020, 5:06 p.m. EST

Given a problem large enough to benefit from cluster computing, then the benefit for time-dependent study is gained by the benefit of each separate time step. (This is essentially the same operation that takes place at end of each iteration of a steady-state problem, BTW.) COMSOL will need to gather up all the results from the split problem over the cluster nodes at the end of each iteration/time step before proceeding to split the work back up to the next iteration/time step.

A more interesting question to me is how many nodes can COMSOL effectively use on a large problem before the overhead required to scatter-gather the data overcomes the benefit in the work splitting of the problem over the nodes ? I still have not seen a study done by a COMSOL user that shows this. In my pre-retirement days, we were limited to 16 nodes; each node composed of large memory and processing nodes. There was no indication of slowdown in using all 16 nodes. The question is what about 100, 1000, or more nodes ?

In the paper/presentation linked below, presented at the 2010 COMSOL Conference in Boston, I showed representative speedup for COMSOL-4.0 when distributed parallel processing was first released. There have been many improvements in the code since that time, so the speedup should be much better now in the present version 5.5.

https://www.comsol.com/paper/exploiting-new-features-of-comsol-version-4-on-conjugate-heat-transfer-problems-7970

-------------------
James D. Freels, Ph.D., P.E.
Given a problem large enough to benefit from cluster computing, then the benefit for time-dependent study is gained by the benefit of each separate time step. (This is essentially the same operation that takes place at end of each iteration of a steady-state problem, BTW.) COMSOL will need to gather up all the results from the split problem over the cluster nodes at the end of each iteration/time step before proceeding to split the work back up to the next iteration/time step. A more interesting question to me is how many nodes can COMSOL effectively use on a large problem before the overhead required to scatter-gather the data overcomes the benefit in the work splitting of the problem over the nodes ? I still have not seen a study done by a COMSOL user that shows this. In my pre-retirement days, we were limited to 16 nodes; each node composed of large memory and processing nodes. There was no indication of slowdown in using all 16 nodes. The question is what about 100, 1000, or more nodes ? In the paper/presentation linked below, presented at the 2010 COMSOL Conference in Boston, I showed representative speedup for COMSOL-4.0 when distributed parallel processing was first released. There have been many improvements in the code since that time, so the speedup should be much better now in the present version 5.5. https://www.comsol.com/paper/exploiting-new-features-of-comsol-version-4-on-conjugate-heat-transfer-problems-7970

Note that while COMSOL employees may participate in the discussion forum, COMSOL® software users who are on-subscription should submit their questions via the Support Center for a more comprehensive response from the Technical Support team.

Suggested Content