As configured, the SLURM scheduler allocates 2G per core requested (where G is SLURM’s abbreviation for GB).
It is our policy that a user requests a proportional amount of memory corresponding to the number of cores requested.
E.g. The nodes on the public queue have 256G each for all 64 CPU cores. If you request 12 cores you can calculate 256/64*12 to obtain the proportional value for your request.
E.g. large-men has 2048G and 64 cores. That’s 32GB per core.
It is important to know each node type’s specifications to accurately calculate this value. Reserving too much memory can render the rest of the node, idle CPUs and GPUs unusable.
Another important node is that the memory specification is not the same as total available memory. Each node of the public queue has 250G of memory available to request.
If you request all of a node, e.g. all 64 cores of a public node, you can request all of the memory with --mem=0. Only do this if you have allocated all of the CPUs