MPI simulations

[Ommair Ishaque]

Hi,

I just to know if I want to run my simulations on single node with multiple cores, do i have compile or install openmm in different way? If not is there a way run simulations on multiple nodes?

[Peter Eastman]

No need. If you use the CPU platform, it will automatically create enough threads to use all available cores.

[Thomas Evangelidis]

Greetings,

I am writing in this thread rather than creating a new one because I encounter a problem with multithreading. I just did a fresh compilation on my PC using the defaults parameters (attached) but when I try to minimize a system on the CPUs I see that only one core is working. The same happens even when I "export OPENMM_CPU_THREADS=12". This is in contrast to what was happening before when I also used the defaults during compilation. I have attached my CMakeCache.txt file. Is there any parameter I should change to get the multi-threading mode running properly?

Thanks in advance.
Thomas

[Peter Eastman]

What platform are you using?

[Thomas Evangelidis]

I am on Ubuntu 18.04 64-bit, CUDA-10. For energy minimization I ALWAYS use the CPU.

[Peter Eastman]

I mean which OpenMM platform, not which OS. What does context.getPlatform().getName() return?

[Thomas Evangelidis]

As I state above, I do the minimization on CPU platform, this is also returned by the command line you wrote. I have compiled from source the dev version on both my PC and my laptop. On my laptop the CPU minimization runs on multiple threads but on my PC on just one. The only difference I can notice is that on my laptop I also have OpenCL libraries installed (and the respective variables 'ON' inside my makefile) but on my PC not. Do you think this is the reason why?

[Peter Eastman]

If you're using the CPU platform, it makes no difference whether the OpenCL platform also happens to be present. It's not getting used.

If 1) you're using the CPU platform, and 2) your processor has multiple cores, and 3) you haven't set OPENMM_CPU_THREADS=1 to force it to only use one thread, then it should be using multiple threads. If it really isn't, that suggests one of the above assumption is wrong. Also make sure nothing else is limiting how many cores it can use (such as OS level process quotas).

[Thomas Evangelidis]

OK, this doesn't help really. I do both use the CPU platform and OPENMM_CPU_THREADS=12.

Anyway, the reason I do the minimization on the CPU is becasue with CUDA mixed precission I get structures with clashes and when I launch MD I get NaN. Now I changed that and do the minimization on CUDA with double precision. So far I didn't have a problem and it's remarkably faster than on 12 CPu cores. Do you think that CPU and GPU double precision are equally credible?

[Peter Eastman]

Do you think that CPU and GPU double precision are equally credible?

It isn't really a question of being credible, since there's no single right answer. The minimizer just searches for any nearby conformation where all the forces are sufficiently close to zero. The two platforms won't necessarily end with the same conformation (and for that matter, running it twice with the same platform may yield two different conformations), but that's ok. As long as it finds something that satisfies your convergence criterion, that's "correct".