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.

Seriously needing speed

Please login with a confirmed email address before reporting spam

We’re going to try this again. About 7 months ago we put two months of effort and a lot of money into a top-of-the-line four-socket machine trying to get more speed than we were getting from a cheap single-socket i7-3930K (6 cores, 3.2 GHz, DDR3, 12MB cache, ~$600) on a large RF problem (~5000 boundaries, ~9M dofs), only to learn the hard way that COMSOL RF wasn’t able to take advantage of modern high-end computers.

We’ll probably buy at least two new computers.

For a low-budget initial experiment, I think we’ll get an i7-5960X (8 cores, 3 GHz, DDR4, 20MB cache, 64 GB max, $1060). For some problems we’ll need a lot more memory than this processor supports, but the point of this experiment is to guide us in buying a high-end machine. If it’s no faster than our current 6-core machine, it tells us that clock speed is about all that matters (notwithstanding what one sees on various COMSOL web pages). (It seems our friends at COMSOL can’t provide any benchmark data, or valid recommendations. We’re definitely not going with more than a single-socket machine. Our experience and that reported by others showed COMSOL is slower on multi-socket boards than on single-socket boards of similar clock speed. If they disagree, they should show benchmark data supporting their claims.)

For a high-end machine, I suppose we’ll go with either the E7-8891-v3 (10 cores, 2.8 GHz, DDR4, 45MB cache, ~$6900) or the E7-8890-v3 (18 cores, 2.5 GHz, DDR4, 45MB cache, $7200), depending on what we learn from the above low-budget experiment.

We’ll post what we learn. (The last time we posted what we learned, the moderator took it down immediately.) We’re hoping that maybe if we keep bugging them, they’ll eventually appreciate that they really need to figure out how to take advantage of modern high-end computers on large problems.

Yes, we do see that COMSOL RF probably does utilize additional cores effectively when it finally gets into the iterative solver, but that’s only a small fraction of the total run time, even with just 6 cores; and most of the time the engineer is sitting waiting on COMSOL (finalizing, meshing, saving, opening, …) it’s only using one core and one memory channel, no matter how many there are. (We put weeks into trying everything we were told to try to get parallelization to work, and never saw any speed up, even with distributed parametric sweeps, on a 4-socket high-end machine.)

Does anyone out there have benchmark data (or even anecdotal indications) comparing some run times and productivity experience with large problems (preferably RF or CFD) on different modern high-end computers?

David

5 Replies Last Post Jul 12, 2017, 4:03 p.m. EDT

Please login with a confirmed email address before reporting spam

Posted: 9 years ago Sep 8, 2015, 3:09 a.m. EDT
Hi,

Speeding is also my problem and I found from some colleague that they have also the same issue.

I am trying to simulate subsurface flow for a big domain (more than 10 ha in 3D), my experience working with COMSOL: we have license just for one pc- which is our workstation (3.19 GHz Intel(R) Xeon cpu, 48 GB RAM, 16 cores(based on the number of boxes in task manager)).

I used with trial version that I installed in my mac(2.4 GHz Intel core i5, 8GB RAM 1600 MHz DDR3), and some times I tried to run the two same model starting at the same time, and I didn't see big difference in time of simulation.

Hope it helps.
Sepideh
Hi, Speeding is also my problem and I found from some colleague that they have also the same issue. I am trying to simulate subsurface flow for a big domain (more than 10 ha in 3D), my experience working with COMSOL: we have license just for one pc- which is our workstation (3.19 GHz Intel(R) Xeon cpu, 48 GB RAM, 16 cores(based on the number of boxes in task manager)). I used with trial version that I installed in my mac(2.4 GHz Intel core i5, 8GB RAM 1600 MHz DDR3), and some times I tried to run the two same model starting at the same time, and I didn't see big difference in time of simulation. Hope it helps. Sepideh

Edgar J. Kaiser Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 9 years ago Sep 8, 2015, 5:23 a.m. EDT
Hi David,

maybe you would like to provide a benchmark model, so some users can run it and report about their machine's performance. I would recommend something that fits well into a 64 GB machine at first. It might comprise a scaling parameter for the mesh that allows to size it up for testing on larger machines. I can run RF, but not CFD.

Cheers
Edgar

--
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
Hi David, maybe you would like to provide a benchmark model, so some users can run it and report about their machine's performance. I would recommend something that fits well into a 64 GB machine at first. It might comprise a scaling parameter for the mesh that allows to size it up for testing on larger machines. I can run RF, but not CFD. Cheers Edgar -- Edgar J. Kaiser emPhys Physical Technology http://www.emphys.com

Please login with a confirmed email address before reporting spam

Posted: 9 years ago Sep 8, 2015, 10:32 a.m. EDT
Edgar,
Yes, this has been suggested several times, but all of our work has been extremely proprietary - until now. I'm working on an RF case now that I'd planned to write-up and post in a month or so that has relatively simple geometry but requires at least 2M dof's and can easily scale to 10 times that. So I'll make this a higher priority - should have it within a week. It will be of interest to all COMSOL users doing MRI and NMR, and to many others doing RF involving multi-resonant circuits with lumped elements. It will make a nice demo, tutorial, and benchmark - much more complex than most tutorials, with many parameters, probes, functions, lumped elements, materials, ports, etc.

Cheers,
David
Edgar, Yes, this has been suggested several times, but all of our work has been extremely proprietary - until now. I'm working on an RF case now that I'd planned to write-up and post in a month or so that has relatively simple geometry but requires at least 2M dof's and can easily scale to 10 times that. So I'll make this a higher priority - should have it within a week. It will be of interest to all COMSOL users doing MRI and NMR, and to many others doing RF involving multi-resonant circuits with lumped elements. It will make a nice demo, tutorial, and benchmark - much more complex than most tutorials, with many parameters, probes, functions, lumped elements, materials, ports, etc. Cheers, David

Edgar J. Kaiser Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 9 years ago Sep 8, 2015, 4:34 p.m. EDT
David,

well I am looking forward to this. However, to be honest, a very complex model is probably not the best benchmark. The benchmark should be quite simple and just challenge the machine, not the user. The simpler it is the better the reproducibility will be.

Cheers
Edgar

--
Edgar J. Kaiser
emPhys Physical Technology
www.emphys.com
David, well I am looking forward to this. However, to be honest, a very complex model is probably not the best benchmark. The benchmark should be quite simple and just challenge the machine, not the user. The simpler it is the better the reproducibility will be. Cheers Edgar -- Edgar J. Kaiser emPhys Physical Technology http://www.emphys.com

Josh Thomas Certified Consultant

Please login with a confirmed email address before reporting spam

Posted: 7 years ago Jul 12, 2017, 4:03 p.m. EDT
Updated: 7 years ago Jul 12, 2017, 4:05 p.m. EDT
David and all-

We also work with models with a large number of geometric entities (~4k domains, ~32k boundaries) and need to do Boolean and other geometric operations.

A couple years ago I came up with a non-proprietary benchmarking model which is simply a 16 X 16 X 16 array of spheres inside of a block. The time to generate and finalize and manipulate this test case is comparable to the real scenario of interest to us (and may be for you, too, David).

I have uploaded a graphic showing a comparison of this benchmarking example between v5.2 and v5.3 on my machine (specs in excel file). I have also uploaded an excel file with the steps to perform this benchmarking example yourself on your system. The excel file includes the time breakdown as well as memory usage and file size.

Great news! In summary--v5.3 released earlier this year--shows almost *3X* improvement of the time to build and manipulate this test geometry.

Many thanks to the COMSOL team!

--
Best regards,
Josh Thomas
AltaSim Technologies
David and all- We also work with models with a large number of geometric entities (~4k domains, ~32k boundaries) and need to do Boolean and other geometric operations. A couple years ago I came up with a non-proprietary benchmarking model which is simply a 16 X 16 X 16 array of spheres inside of a block. The time to generate and finalize and manipulate this test case is comparable to the real scenario of interest to us (and may be for you, too, David). I have uploaded a graphic showing a comparison of this benchmarking example between v5.2 and v5.3 on my machine (specs in excel file). I have also uploaded an excel file with the steps to perform this benchmarking example yourself on your system. The excel file includes the time breakdown as well as memory usage and file size. Great news! In summary--v5.3 released earlier this year--shows almost *3X* improvement of the time to build and manipulate this test geometry. Many thanks to the COMSOL team! -- Best regards, Josh Thomas AltaSim Technologies

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.