Running Mumax3 on multi-GPU

Skip to first unread message

Charudatta Phatak

unread,
Sep 22, 2021, 12:00:07 PM9/22/21
to mumax2
Hi,

I was wondering if the current version of mumax (mumax3.10_linux_cuda11.0) supports running on multiple GPUs? I am trying to run it on the new Nvidia A100 node which has 8 GPUs associated with it. Does mumax run in a distributed mode automatically or are there any specific flags that need to be added?

Thank you for you help in advance!

cheers,
-cd

Josh Lauzier

unread,
Sep 24, 2021, 3:25:50 AM9/24/21
to mumax2
Hello,

Yes, mumax comes with the ability to distribute jobs (you can only use 1 GPU per simulation). You can specify a particular gpu with the gpu flag, as well.

Some additional capability is given with the mumax3-server package for clusters, which may be useful for you. It has the ability to schedule/queue jobs etc.

Best,
Josh L

Felipe Garcia

unread,
Sep 24, 2021, 6:11:37 AM9/24/21
to mumax2
Hi Charudatta Phatak,

One can use each GPU for a different calculation. I think it is not possible to do real distributed calculations of a single script. As Josh Lauzier mentioned it is possible to do a kind of batch system with mumax3-server. The other option is to use
mumax3 script1.txt script2.txt ... scriptn.txt

This will automatically distribute the script between the GPUs until all the scripts are finalized. It will occupy all the available GPU even if there is another mumax3 script running.

Best regards,
Felipe

--
You received this message because you are subscribed to the Google Groups "mumax2" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mumax2+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mumax2/01dfd7be-5f1b-4c2d-bae7-5baaf585260cn%40googlegroups.com.

Charudatta Phatak

unread,
Sep 24, 2021, 12:43:32 PM9/24/21
to mumax2
Thank you, Felipe and Josh L for your answers. That is quite helpful. 

cheers,
-cd

Myles J

unread,
Oct 2, 2021, 11:12:46 AM10/2/21
to mumax2
Hi CD:

How large are your problems (in terms of grid-size) and how long does it take to run on one gpu (time to solution)?  A single gpu can handle a lot of work (almost 100x of a regular cpu core) - so unless you have a very large problem, running on multiple gpu cores might not be necessary for a single problem.

Myles
Reply all
Reply to author
Forward
0 new messages