Dear LAMMPS Developers and Users,
I was trying to optimize my simulation runs using “numa” style of grid type with processors command but I am getting error for some number of cores like 32 cores, but is strangely working for 24 or 48 cores. The error I am getting is:
ERROR: Could not create numa grid of processors (…/procmap.cpp:189)
I am running the simulation on a system with 24 cores per node. Is it because number of cores used for simulation is not a multiple of number of cores/node (i.e. 24 here)? Can something be done to work around this but still use numa style?
I would like to apologize since I had failed to check documentation properly before posting here. It was because number of cores used for simulation was not a multiple of number of cores/node (i.e. 24 here). But can something be done to work around this but still use numa style to reduce off node communication?
A more fundamental Q is if you are using a machine with
24 cores/node, should you ever request an MPI allocation that
is not a multiple of 24. Most big machines only allow
you to request a number of nodes, not a fraction of
cores on different nodes. NUMA mapping does not
really make sense in the latter scenario.