nss_slurm

nss_slurm is an optional NSS plugin that can permit passwd and group resolution for a job on the compute node to be serviced through the local slurmstepd process, rather than through some alternate network-based service such as LDAP, SSSD, or NSLCD.

When enabled on the cluster, for each job, the job's user will have their full struct passwd info — username, uid, primary gid, gecos info, home directory, and shell — securely sent as part of each step launch, and cached within the slurmstepd process. This info will then be provided to any process launched by that step through the getpwuid()/getpwnam()/getpwent() system calls.

For group information — from the getgrgid()/getgrnam()/getgrent() system calls —, an abbreviated view of struct group will be provided. Within a given process, the response will include only those groups that the user belongs to, but with only the user themselves listed as a member. The full list of group members is not provided.

Installation

Source:

In your Slurm build directory, navigate to contribs/nss_slurm/ and run: 

make && make install
This will install libnss_slurm.so.2 alongside your other Slurm library files in your install path.

Depending on your Linux distribution, you will likely need to symlink this to the directory which includes your other NSS plugins to enable it. On Debian/Ubuntu, /lib/x86_64-linux-gnu is recommended, and for RHEL-based distributions /usr/lib64 is recommended. If in doubt, a command such as find /lib /usr/ -name 'libnss*' should help.

Setup

The slurmctld must be configured to lookup and send the appropriate passwd and group details as part of the launch credential. This is handled by setting LaunchParameters=enable_nss_slurm in slurm.conf and restarting slurmctld.

Once enabled, the scontrol getent command can be used on a compute node to print all passwd and group info associated with job steps on that node. As an example: 

tim@node0001:~$ scontrol getent node0001
JobId=1268.Extern:
User:
tim:x:1000:1000:Tim Wickberg:/home/tim:/bin/bash
Groups:
tim:x:1000:tim
projecta:x:1001:tim

JobId=1268.0:
User:
tim:x:1000:1000:Tim Wickberg:/home/tim:/bin/bash
Groups:
tim:x:1000:tim
projecta:x:1001:tim

NSS Slurm Configuration

nss_slurm has an optional configuration file — /etc/nss_slurm.conf. This configuration file is only needed if:

  • The node's hostname does not match the NodeName, in which case you must explicitly set the NodeName option.
  • The SlurmdSpoolDir does not match Slurm's default location of /var/spool/slurmd, in which case it must be provided as well.

NodeName and SlurmdSpoolDir are the only configuration options supported at this time.

Initial Testing

Before enabling NSS Slurm directly on the node, you should use the -s slurm option to getent within a newly launched job step to verify that the rest of the setup has been completed successfully. The -s option to getent allows it to query a specific database — even if it has not been enabled by default through the system's nsswitch.conf. Note that nss_slurm only responds to requests from processes within the job step itself — you must launch the getent command within a job step to see any data returned.

As an example of a successful query: 

tim@blackhole:~$ srun getent -s slurm passwd
tim:x:1000:1000:Tim Wickberg:/home/tim:/bin/bash
tim@blackhole:~$ srun getent -s slurm group
tim:x:1000:tim
projecta:x:1001:tim

NSS Configuration

Enabling nss_slurm is as simple as adding slurm to the passwd and group database in /etc/nsswitch.conf. It is recommended that slurm is listed first, as the order (from left to right) determines the sequence in which the NSS databases will be queried, and this ensures Slurm handles the request if able before submitting the query to other sources.

Once enabled, test it by launching getent queries such as: 

tim@blackhole:~$ srun getent passwd tim
tim:x:1000:1000:Tim Wickberg:/home/tim:/bin/bash
tim@blackhole:~$ srun getent group projecta
projecta:x:1001:tim

Limitations

nss_slurm will only return results for processes within a given job step. It will not return any results for processes outside of these steps, such as system monitoring, node health checks, prolog or epilog scripts, and related node system processes.

nss_slurm is not meant as a full replacement for network directory services such as LDAP, but as a way to remove load from those systems to improve the performance of large-scale job launches. It accomplishes this by removing the "thundering-herd" issue should all tasks of a large job make simultaneous lookup requests — generally for info related to the user themselves, which is the only information nss_slurm will be able to provide — and overwhelm the underlying directory services.

Last modified 2 July 2019