Am 05.12.2015 um 19:02 schrieb Dj Merrill: > On 12/5/2015 11:44 AM, Reuti wrote: >> Yes, it will be multiplied too. But how does the job behave in case you want >> to run 8 processes with an MPI job, but only one GPU is installed. Will only >> one process/thread access it in a serial step? Doesn't it imply it to have a >> number of GPUs installed according to the number of slots you request? > > You are correct, of course, for standard MPI jobs. Asking for pe=8 is in > effect 8 separate jobs that just happen to talk to each other, and only one > of those jobs can effectively use the GPU. > > I was thinking along the lines of custom MPI or SMP jobs that detect whether > the GPU is in use (via CUDA tools or similar), and use the GPU if available, > otherwise just use cpus. That would allow one of the 8 processes to use the > GPU, and the other 7 to use cpu only.
I would assume that this will need sophisticated programming, as the processes/threads are no longer running at the same speed then. Replacing just one CPU by a GPU will still make the other processes not faster and the fast one will have to wait for the other ones to finish too. I wonder about the variance in mathematical precision whether the CPU or a GPU is used too. > This might be a fringe use case, though. > >> >> Define the gpu complex as consumable JOB, then it won't be multiplied. > > Ah, yes, that makes sense. I have it set as YES, not JOB. Thanks for the > pointer! > > $ qconf -sc | grep gpu > gpu gpu INT <= YES YES 0 0 > > Should be: > > # qconf -sc | grep gpu > gpu gpu INT <= YES JOB 0 0 Yep. -- Reuti > I will poke at this on Monday and hopefully not mess things up too badly... > :-) > > Rajil, does this help you as well? I think this answers the same question(s) > that I had. > > -Dj > > > _______________________________________________ > users mailing list > [email protected] > https://gridengine.org/mailman/listinfo/users _______________________________________________ users mailing list [email protected] https://gridengine.org/mailman/listinfo/users
