Page 1 of 1

Massive parallelization

Posted: Fri Mar 18, 2016 1:56 am
by ramidas
Hello,

Does openMM simulations can be run on multiple nodes in computing cluster? I mean, on multiple nodes with multiple cores. With openMPI or something like that.

Re: Massive parallelization

Posted: Fri Mar 18, 2016 9:50 am
by peastman
A single simulation can only be run on a single node. Of course, there are other ways of taking advantage of multiple nodes that are often more efficient. For example, you could do replica exchange (a different replica on each node), or run many independent simulations to get better sampling.

Peter

Re: Massive parallelization

Posted: Thu Mar 24, 2016 11:17 am
by gpantel
This is something I am also curious about - running one replica per node would be fine for some replica exchange ideas I've been playing with recently. How would someone go about running REMD over many nodes with OpenMM? How would the implementation have to be built?

Thanks you for any info,
George

Re: Massive parallelization

Posted: Thu Mar 24, 2016 11:28 am
by peastman
There's already an implementation of it:

https://github.com/choderalab/repex

Peter

Re: Massive parallelization

Posted: Thu Mar 24, 2016 11:56 am
by gpantel
Yes, I know of John Chodera's repex and yank - I mean if I were to want to do this myself. How would someone set different replicas to run on specific nodes?

Re: Massive parallelization

Posted: Thu Mar 24, 2016 12:02 pm
by jswails1
I would use MPI, personally. But that's precisely what John Chodera's library does.

So I guess the best answer I can think of is to do what repex does. Which is to say, just use repex (or implement the same thing in C++ if you want to do that and can reproduce the rest of the required machinery from the app layer in C++). The code is free to look at, use, modify, and reuse.

Re: Massive parallelization

Posted: Thu Mar 24, 2016 12:15 pm
by gpantel
Ah, OK, I will just go scour his code, then. Thanks for the advice.