Skip to content

Latest commit

 

History

History
91 lines (73 loc) · 4.14 KB

README.md

File metadata and controls

91 lines (73 loc) · 4.14 KB
Raw

Slurm node scripts

Some convenient scripts for working with nodes (or lists of nodes):

  • Drain a node-list: sdrain node-list "Reason".
  • Resume a node-list: sresume node-list.
  • Reboot and resume a node-list: sreboot node-list.
  • Show node status: shownode <node-list>.
  • Show node power values: showpower < -w node-list | -p partition(s) | -a | -h > [ -S sorting-variable ] -s.
  • Check dead nodes with a ping: alive (may be run from crontab every 10 minutes)
  • Show Nvidia GPU power values: showpower_nvidia < -w node-list | -p partition(s) | -a | -h > -s
  • Do a ps process status on a node-list, but exclude system processes: psnode [-c columns | -h] node-list.
  • Print Slurm version on a node-list: sversion node-list. Requires ClusterShell.
  • Check consistency of /etc/slurm/topology.conf with node-list in /etc/slurm/slurm.conf: checktopology
  • Compute node OS and firmware updates using the update.sh script.

Usage

Copy these scripts to /usr/local/bin/. If necessary configure the variables in the script.

Example output from psnode:

Node d064 information:
NODELIST   PARTITION  CPUS  CPU_LOAD    S:C:T  MEMORY       STATE REASON              
d064       xeon8*        8      2.01    2:4:1   47000       mixed none                
d064       xeon8_48      8      2.01    2:4:1   47000       mixed none                
Jobid list: 3381322 3380373
Node d064 user processes:
  PID NLWP S USER      STARTED     TIME %CPU   RSS COMMAND
19984    1 S user1      Jan 19 00:00:00  0.0  2224 /bin/bash -l /var/spool/slurmd/job3380373/slurm_s
20092    1 S user1      Jan 19 00:00:00  0.0  1368 /bin/bash -l /var/spool/slurmd/job3380373/slurm_s
20094    3 R user1      Jan 19 1-06:25:18 99.9 256676 python3 /home/user1/wlda/atomic_bench
20096    5 S user1      Jan 19 00:00:01  0.0 15136 orted --hnp --set-sid --report-uri 8 --singleton-
27564    1 S user1    22:42:23 00:00:00  0.0  2228 /bin/bash -l /var/spool/slurmd/job3381322/slurm_s
27673    1 S user1    22:42:27 00:00:00  0.0  1372 /bin/bash -l /var/spool/slurmd/job3381322/slurm_s
27675    3 R user1    22:42:27 10:11:58 99.9 242464 python3 /home/user1/wlda/atomic_benchma
27676    5 S user1    22:42:27 00:00:00  0.0 15132 orted --hnp --set-sid --report-uri 8 --singleton-
Total: 8 processes and 20 threads

Compute node OS and firmware updates

This procedure requires ClusterShell.

Assume that you want to update OS and firmware on a specific set of nodes defined as <node-list>. It is recommended to update entire partitions, or the entire cluster, at a time in order to avoid having inconsistent node states in the partitions.

First configure the update.sh script so that it will perform the required OS and firmware updates for your specific partitions.

Then copy the update.sh file to the compute nodes:

clush -bw <node-list> --copy update.sh --dest /root/

On the compute nodes append this crontab entry:

clush -bw <node-list> 'echo "@reboot root /bin/bash /root/update.sh" >> /etc/crontab'

Then set the nodes to make an automatic reboot (via Slurm) as soon as they become idle (ASAP, see the scontrol manual page) and change the node state to DOWN with:

scontrol reboot ASAP nextstate=DOWN reason=UPDATE <node-list>

You can now check nodes regularly (a few times per day) as the rolling updates proceed. List the DOWN nodes with sinfo -lR.

Check the status of the DOWN nodes. For example, you may check the running kernel and the BMC version, and use NHC:

clush -bw@slurmstate:down 'uname -r; nhc; dmidecode -s bios-version'

When some nodes have been updated and tested successfully, you could resume these nodes by:

scontrol update nodename=<nodes that have completed updating> state=resume

Resuming the node is actually accomplished at the end of the update.sh script by these lines:

scontrol reboot nextstate=resume `hostname -s`