Slurm has an optional module called nss_slurm (Slurm Workload Manager - nss_slurm). On our larger systems, the load on LDAP can get quite heavy and the scheduler is well situated to cache the necessary user and group data locally to avoid that. I looked through the docs and forums and did not see anything equivalent, but wanted to ask in case I missed it.
Assuming it is not there, if anyone has any thoughts on how / where to plug this in, we would be happy to hear about it.
We just installed/run nscd on our PBS servers (and doing so cut our server start up time significantly). I’m a little confused about what nss_ldap would do that nscd doesn’t. (To be fair I don’t do much with ldap).
Thanks for the input. I will run the idea of ncsd by our admins and see what they think.
I talked to our admins. We are currently using sssd, which is, as I understand it, a successor to ncsd. It works, but there are issues, like the negative cache not clearing properly, and we need to run additional LDAP replicas because of the size our machines. So we have a workable solution, but they believe the nss functionality in the scheduler would be a cleaner solution, and would give us parity with Slurm, which is run at many of the other DOE facilities.
I don’t know the ins and outs of when we talk to the system for user and group info, but I do know once a job is submitted, the username and group name are stored in attributes. I don’t think we talk to the system again until the mom does when the job runs.