We get problems with running out of memory if we have too many NWChem tasks running on a MOM node.
The issue is likely that all of the stdout and a few output files from each run are stored in memory. Perhaps we could modify QCEngine to route the stdout to disk and read data from a mmaped file?
Or, could we at least place the processes that launch MPI tasks/read from stdout on their own node with more memory. Ideally, we could move the whole "thinker" to occupy this node.
WardLT
commented
3 years ago
We get problems with running out of memory if we have too many NWChem tasks running on a MOM node.
The issue is likely that all of the stdout and a few output files from each run are stored in memory. Perhaps we could modify QCEngine to route the stdout to disk and read data from a mmaped file?
Or, could we at least place the processes that launch MPI tasks/read from stdout on their own node with more memory. Ideally, we could move the whole "thinker" to occupy this node.