We've actually been patching the Slurm cgroup plug-in to enable configurable per-node and per-partition swap settings. E.g. on node X with 64 cores a job gets (N_core,job / N_core,tot)*8 GiB added to the physical RAM limit, where 8 GiB is some fraction of the total swap available. It's still not technically a swap limit per se, because it entails the sum of swap + physical RAM usage, but it has kept our nodes from getting starved out thanks to heavy swapping and can scale with job size, etc.
/*! @signature Jeffrey Frey, Ph.D @email f...@udel.edu @source iPhone */ > On Aug 17, 2022, at 04:13, Hermann Schwärzler > <hermann.schwaerz...@uibk.ac.at> wrote: > > Hi Eg. > > if you are using cgroups (as you do if I read your other post correctly) > these two lines in your cgroup.conf should do the trick: > > ConstrainSwapSpace=yes > AllowedSwapSpace=0 > > Regards, > Hermann > > PS: BTW we are planning to *not* use this setting as right now we are looking > into allowing jobs to swap/page when they have exhausted their requested > memory. > Until now our impression is that this has an impact on the respective job > only and not on the whole compute node. But maybe we are wrong... > > >> On 8/11/22 10:51 PM, Eg. Bo. wrote: >> Hello, >> could anybody share a decent configuration which prohibits swap usage for >> any job? Right now, I'd think my configuration (see my previous post in >> regards to RealMemory configuration) could end up in swapping which is not >> intended at all for compute nodes. >> Thanks&Best >> Eg. >
