Hello everybody, we are in a discussion about using slurm and I have some questions. If there is a more appropriate mailing list to ask these please let me know. Also, if this has been asked before I would appreciate any links to the previous discussion or documentation, I was not able to find any.
We would like to make sharing of nodes between jobs of different users an option. With our current system (maui/troque) resource usage, i.e. primarily CPU time, does not appear to be billed "correctly" in this configuration. For example, if a job requests 4 out of 20 cores on a node but sets NACCESSPOLICY:SINGLEJOB as option to qsub I have been told it still is billed only for the 4 cores requested and not for the 20 cores blocked. The same appears to be true if we set the option directly in maui.cfg, which is even worse considering the implications. The situation could probably be mitigated by forcing the user to set appropriate memory resource requests at job submission time, but still this usage computation is not what we would consider truly transparent, especially considering the somewhat lacking documentation of "FSPOLICY DEDICATEDPES". Now, after that background I hope my questions make sense in the slurm context. If I understand the slurm documentation correctly the usual configuration in slurm would be to allocate resources (i.e. whole nodes) exclusively to a single job only. Setting select/cons_res (and subsequent configuration of consumables) as well as OverSubscribe Yes would in principle enable a user to decide whether to allow sharing of resources by setting "--share/-s" or "--exclusive" as options to sbatch. Is this correct so far ? How is the resource usage calculated then ? Assume a job requests only for example 4 out of 20 cores on a node, but sets "--exclusive" effectively blocking the whole node (e.g. if user does not want to think about memory). Would the CPU usage of that job then be computed based on 4 core or the 20 cores it basically prevents from being used by other users ? We would like to simply track cpus in TRES in this scenario. Conversely, if slurm defaults are used and one node is allocated exclusively to one job then the TRES to use is just "node" ? Of course also suggestions/pointers how to better achieve this in slurm would be appreciated. Thank you very much in advance ! Best Regards Christof -- Dr. rer. nat. Christof Köhler email: [email protected] Universitaet Bremen/ BCCMS phone: +49-(0)421-218-62334 Am Fallturm 1/ TAB/ Raum 3.12 fax: +49-(0)421-218-62770 28359 Bremen PGP: http://www.bccms.uni-bremen.de/cms/people/c_koehler/
signature.asc
Description: Digital signature
