jrhaberstroh
commented
11 years ago http://www.gromacs.org/Documentation/Acceleration_and_parallelization
Locking threads to physical cores By default thread-MPI and the OpenMP parallelization, active when the verlet cut-off scheme is used, fill up all cores in the machine. When all cores used, mdrun will pin the threads to specific cores, if this hasn't already been done by MPI or OpenMP. If you want optimal performance when not using all cores, you need to use mdrun -pin on. If you want to run multiple jobs on the same compute node, you need to limit the number of cores used and if you want good performance, pin different jobs to different cores. The mdrun option -nt sets the total number of threads for an mdrun job. The -pinoffset option sets a pinning offset, counted in logical numbers of cores. For example, running 2 mdrun jobs on an Intel CPU with 6 physical cores with hyper-threading (supporting 2 threads) can be achieved with:
mdrun -nt 6 -pin on -pinoffset 0 mdrun -nt 6 -pin on -pinoffset 3
No implementation of cluster_traj_all for hopper, partly because multitenancy is unavailable, and 24 cores is excessive for most jobs. Figuring out how to do multi-tenancy jobs on hopper would be a boon.