Ticket 5380

Hi Ben,

looking at the code, I see this:

        if (opts.node_name)
                cgroup_suffix = xstrdup_printf("_%s", opts.node_name);

        if (snprintf(uidcg, PATH_MAX, "%s/memory/slurm%s/uid_%u",
                     slurm_cgroup_conf->cgroup_mountpoint, cgroup_suffix, uid)
            >= PATH_MAX)

in the _indeterminate_multiple() function. Meaning if nodename option is specified in the PAM module, then .../slurm%s/... will be substituted with such name to generate the final path. After that there's a call _cgroup_creation_time() to end constructing the path string with the %s/job_%u information. Could you try to add the "nodename" option to the module? Something like this:

account    sufficient    pam_slurm_adopt.so nodename=caldera14

and see if it works?

I'm not sure why the nodename option isn't mentioned in the guide

https://slurm.schedmd.com/pam_slurm_adopt.html

Will discuss that internally and come back to you. Thanks.

Yup, that seems to work. I saw that code as well, but assumed that it should be able to figure out the nodename (unless people were doing something odd with NodeAddr) and that I was missing something. 

It be nice not to have to customize the pam config for every node

Did you by any chance happen to compile with --enable-multiple-slurmd option?

After a few local tests and inspecting the code, I _suspect_ slurmd daemon generates the cgroup path in the form of .../slurm_<NodeName>/... only when the above config option is enabled. Otherwise, it is generated with the form .../slurm/..., which doesn't require the PAM module to be explicitly configured with the nodename=<NodeName> option, which is what you want.

This reasoning is backed up by this code:

https://github.com/SchedMD/slurm/blob/51606c09536df28373dbabe530fed1c7d4d46426/src/slurmd/common/slurmd_cgroup.c#L322

Although something might be escaping to this theory. In the last config copy we have from your site, your node definitions have NodeName != NodeAddr, although that doesn't seem to be enough for slurmd to create the path with .../slurmd_<NodeName>/... format. So I am inclined to suspect you compiled with multiple slurmd option. Would you mind to check it?

You can know whether you compiled with such option by inspecting your config.log in your build directory, like this:

alex@ibiza:~/slurm/17.11/ibiza/slurm$ grep MULTIPLE_SLURMD config.log
| #define MULTIPLE_SLURMD 1
#define MULTIPLE_SLURMD 1
alex@ibiza:~/slurm/17.11/ibiza/slurm$

If it is not enabled grep doesn't report any line. Thanks.

(In reply to Alejandro Sanchez from comment #7)
> Did you by any chance happen to compile with --enable-multiple-slurmd option?

Looks like yes. I have a coworker who has decided that all the features need to be turned on. I'll go have another talk with him. Thanks for the catch/sorry about the ticket. 

It would be nice if pam_slurm_adopt had a heuristic to handle this, but it's probably enough of an edge case that it might not be worth the effort. 

NodeName is != NodeAddr for other reasons, we're not actually using multiple slurmd for anything.

I think it is not worth the effort either. Thanks for your feedback. We will change the documentation to clarify the cgroup path formats depending upon that option.