Quoting Vsevolod Nikonorov <[email protected]>: > > Hello. > > I have some questions about SLURM's functionality. > > 1. We have a cluster with rather small amount of physical nodes, and > in order to fulfill our users requests more precisely we would like > our partitions to have, for example, 2.5 physical nodes. Since SLURM > can allocate certain amounts of memory on nodes, maybe it possible > what we want?
SLURM partitions must contain entire nodes, but you could execute two slurmd daemons on a compute node, each configured as a different logical node with some of the CPUs of the physical node: http://www.schedmd.com/slurmdocs/faq.html#multi_slurmd > 2. Is it possible to dynamically increase/decrease partition's > resources? For example, we have two partitions with three-hour time > limit for two workgroups which work simultaneously. Of course, > sometimes there will be two in dependent queues of tasks. But if one > partition has a long queue, while another partition has no queue at > all, we may want to reallocate several nodes from one partition to > another, but, like, say them: "Well, now you are allocated to > partition1, but the moment queue disappears, you put yourself in > DRAIN status and then, once you have done your jobs, you go back to > partition2 and become IDLE". Is something like that possible? Event triggers might be used for this. See "man strigger" or http://www.schedmd.com/slurmdocs/strigger.html > Thank you. > > -- > Vsevolod Nikonorov, > NIKIET > > <[email protected]> >
