Dear AMBER community:
How can I run pH-REMD assigning more than one replica to a GPU? I only have 4 GPUs on a local workstation, but each one has ~24 GB of memory. How do I modify the below command?
mpirun -np 4 $AMBERHOME/bin/pmemd.cuda.MPI -ng 4 -groupfile groupfile -rem 4 -remlog rem.log &> output.log &
I’ve tried increasing the ng flag but the program complains that np must be an integer multiple of ng.
Curiously, when I use the above command. I notice that the expected number of jobs (1 in this case) gets assigned to GPU-0 while the other three replicas get shared by GPU-0 and the other GPUs (1, 2, and 3). If I increase the np and ng flags to exceed the number of GPUs, I see the same behavior: GPU-0 has it’s own set of replicas and shares in the replicas that should belong to the other GPUs. There seems to be peer-to-peer communication between all four GPUs on the workstation. Is this normal behavior for running replica exchange on GPUs? Is there a reason why GPU-0 has to participate in all the calculations? Is there a more efficient way to run these calculations?
Best,
Matthew
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Aug 23 2024 - 05:30:02 PDT
