Known issues

From Alliance Doc
Revision as of 16:59, 18 August 2017 by Nathanw (talk | contribs)
Jump to navigation Jump to search
Other languages:

Report an issue

Shared issues

Scheduler problems

  • The CC Slurm configuration encourages whole-node jobs. When appropriate, users should request whole-node rather than per-core resources. See Job Scheduling - Whole Node Scheduling.
  • CPU and GPU backfill partitions have been created on the Cedar and Graham clusters. If a job is submitted with <24hr runtime, it will be automatically entered into the cluster-wide backfill partition. This partition has low priority, but will allow increased utilization of the cluster by serial jobs.
  • Slurm epilog does not fully clean up processes from ended jobs, especially if the job did not exit normally.
    • This has been greatly improved after the addition of the epilog.clean script, but there are still nodes occasionally marked down for epilog failure.
  • By default, the job receives environment settings from the submitting shell. This can lead to irreproducible results if it's not what you expect. To force the job to run with a fresh-like login environment, you can submit with --export=none or add #SBATCH --export=NONE to your job script.


Quota and filesystem problems

Quota errors on /project filesystem

Users will sometimes see a quota error on their project folders. This may happen when files are owned by a group other than the project group. You can change the group which owns files using the command

Question.png 
[name@server ~]$ chgrp -R <group> <folder>

To see what the value of <group> should be, run the following command:

Question.png 
[name@server ~]$ stat -c %G $HOME/projects/*/

Only the owner of the files can run the chgrp command. To ask us to correct the group owner for many users, write to CC support.

Nearline

Missing symbolic links to project folders

Cedar only

  • Slurm operations will occasionally time out with a message such as Socket timed out on send/recv operation or Unable to contact slurm controller (connect failure). As a temporary workaround, attempt to resubmit your jobs/commands; they should go through in a few seconds.
    • Should be resolved after a VHD migration to a new backend for slurmctl.
  • Some people are getting the message error: Job submit/allocate failed: Invalid account or account/partition combination specified.
    • They need to specify --account=<accounting group> in their job submission.
    • For users with only one account, this should be resolved in the latest version of the job submission script (Nathan Wielenga 2017-08-18).

Graham only

  • /home is on an NFS appliance that does not support ACLs, so setfacl/getfacl doesn't work there.
    • Workaround: use the /project or /scratch filesystems instead.
    • Might be resolved by an update or reconfiguration.
  • Compute nodes cannot access Internet
    • Solution: Request exceptions to be made at CC support describing what you need to access and why.
  • Crontab is not offered on Graham. When attempting the addition of a new item, there is an error during saving:
[rozmanov@gra-login1 ~]$ crontab -e
no crontab for rozmanov - using an empty one
crontab: installing new crontab
/var/spool/cron/#tmp.gra-login1.XXXXKsp8LU: Read-only file system
crontab: edits left in /tmp/crontab.u0ljzU

As crontab works on Cedar, there must be some kind of common approach for CC systems. Clearly, the main issue is how to handle user's crontabs on multiple login nodes; it is not clear however if this should be implemented.

Other issues

  1. Modules don't work for shells other than bash(sh) and tcsh.
    • Workaround: (this appears to work but not tested extensively)
      • source $LMOD_PKG/init/zsh
      • source $LMOD_PKG/init/ksh
Retrieved from "https://docs.alliancecan.ca/mediawiki/index.php?title=Known_issues&oldid=36624"