Imagine a heterogeneous cluster ranging from small four core systems to large 40 core machines and user jobs which can practically scale to any number of local cores itself. This page shows how a job can ask request as many cores as available on a node, or ask for a minimum/maximum range of nodes.
In all cases, the job environment variable OMP_NUM_THREADS will be set to the number of cores assigned to the slot where the job is running (the name of that environment variable is a POSIX/OpenMP standard).
** Question: I’d like at least 3 cpus, up to max of 8 if available
executable = foo.exe # Only interested in slots with 3 to 8 CPUs if they # are static or dynamic. If a slot is partitionable, # make sure it has at least 3 cpus. requirements = Cpus >= 3 && (PartitionableSlot || Cpus <= 8) # Request up to 8 cpus out of a partitionable slot. request_cpus = Cpus > 8 ? 8 : Cpus queue
** Question: I’d like all CPUs currently available in a slot, be it either static or partitionable, and I'd like to be matched with machines that have as many cores available as possible.
executable = foo.exe request_cpus = Cpus rank = Cpus queue
** Question: I'd like all the CPUs on the machine; in other words, I want to be the only job running on the machine. Of course, this job may have to wait a long time to run!
requirements = TotalCpus == Cpus executable = foo.exe request_cpus = TotalCpus queue