Dear AMBER Community,
How can I run multiple replicas for constant pH MD on the same GPU? My system is only ~68K atoms and I have GPUs with ~24 GB of memory, so, if doable, this seems advantageous.
When I set the -ng flag for pmemd.cuda.MPI to be greater than the -np flag for mpirun, AMBER22 complains that np needs to be a multiple of ng. If I set np ≥ ng, the code runs, but it looks like some replicas are distributed among multiple GPUs. For example, if I do:
export CUDA_VISIBLE_DEVICES=0,1,2,3
nohup mpirun -np 8 $AMBERHOME/bin/pmemd.cuda.MPI -ng 8 -groupfile groupfile -rem 4 -remlog rem.log &> output.log &
Two replicas run exclusively on GPU0 and 6 replicas each get shared by GPU0 and one of the other GPUs. This seems non-optimal to me because GPU is handling 8 jobs while the other GPUs only have two jobs each.
What is the right command I should be using to run two replicas on each of the four GPUs?
Best,
Matthew
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sun Aug 18 2024 - 00:30:02 PDT
