Discussion Forum

Ok, I can’t really claim COMSOL-CFD is always 10X faster under Linux than under Windows 10, but I can say large k-w models of compressible non-isothermal high-velocity flows in rotating machinery run more than 10 times faster under Linux than they do under Windows 10 on a computer based on two E5-2697v4’s (Q1’16, 512 GB DDR4/2400, 16 cores in each processor, 40MB cache) on a motherboard using the C612 chip set and with a SSD.

We have been screaming to COMSOL support for more than 2 years that we need someone at COMSOL to take seriously the problem that COMSOL-CFD-Mixer can’t use more than a very small fraction of the available CPU resources in modern high-end multi-core machines under Windows 10, at least for large k-w rotating machinery models (and probably for other types of CFD models too). We have demonstrated this serious deficiency on at least 7 machines and hundreds of widely differing cases, with every imaginable change in solver parameters.

No one really took our problem seriously, though we know many COMSOL users have had the same problem (some have asked us for guidance !).

We noted that several months ago, after a Windows 10 update, several machines which were already terribly slow, started running even slower, by more than another factor or 3 !

Finally, a month ago, Bjorn Sjodin, suggested in a phone call that we try Linux. He was aware of some serious issues using modern multi-core processors under Windows 10. He said some code modules were generated using different compilers for running under Linux than under Windows, and they might be better able to utilize available resources. So, in desperation, we got Linux going on one of our machines (as described above). There was a learning curve. We still have a lot to learn. But we got COMSOL to issue the needed license, and yesterday we got our first results.

10 TIMES FASTER !!

We ran a large model (more than 1M meshcells) under 5.2a because that is what we’ve been using for these large cases – because 5.2a handled our cases involving very high shear stresses in transonic flows better than 5.3a did.

We did not rebuild the model or the mesh, or change any initial conditions or model parameters. We didn’t change any solver parameters. We ran a single instance with everything exactly as what was previously working best under Windows 10 – direct solver, PARDISO, 6 iterations from the same initial conditions, no changes in damping or initial CFL number, or anything else.

A couple months earlier on this same computer under Windows 10, this model took 5.1 hours to run. Now, under Linux, this model runs in 30.1 minutes.

THANK YOU BJORN SJODIN !!

We still have a lot to learn about Linux. We have to get Linux versions of all the other software we need to have at our finger tips on this machine. We will soon install COMSOL 5.3a and get this model going under it (may take a while, as we’ll have to change some meshing and solver parameters…). We expect we’ll again be ecstatic. We’ve previously seen little difference in solver speed between 5.2a and 5.3a on any machine under Windows 10 for the same number of iterations. We expect we’ll see the same under Linux. (We have seen it usually takes more iterations under 5.3a to get to the same convergence, but that’s another, and comparatively minor, issue.)

A few months ago we bought the fastest computer we could justify: Two Xeon gold 6138’s, Q3’17; 20 cores, 2GHz, 3.7GHz turbo, 6 memory channels; 22MB cache; 512GB DDR4/2666. Supermicro X11DAi-N Dual Intel Skylake-SP Xeon WS board. We ran the same CFD test case on this computer, under Windows 10, and it ran almost exactly as slow as on the earlier computer (the one with dual 2697’s) even though this computer runs most standard benchmarks 20-30% faster.

We’re now optimistic that we can switch from Windows 10 to Linux on this Xeon Gold machine and see another 20-30% increase in the speed of COMSOL-CFD. We’ll probably try that at some point, in a month or so, after we’re fully comfortable with Linux. But for now, we’re very happy with a factor of 10 increase in solver speed, and optimistic that we’ll be able to move to 5.3a under Linux and take advantage of its improvements in geometry handling, selections, and meshing, and eventually figure out some remaining solver issues.

I suppose some moderator will soon take this post down because it says COMSOL-CFD-Mixer is not compatible with Windows 10 on modern high-end machines. But I hope not. I hope COMSOL highlights it. There are probably hundreds of users out there (maybe thousands) hoping someone will tell them how to get COMSOL-CFD to run large problems effectively on modern high-end machines.

The answer is (relatively) simple. Switch from Windows 10 to Linux !! Then COMSOL-CFD works !!

David Doty, PhD