Correction/addendum: If the node you want to exclude has RPMS that were
built without NVML autodetection, you probably want that gres.conf to
look like this:
NodeName=a1-10 Name=gpu File=/dev/nvidia0
I'm guessing if it was built without Autodetection, the AutoDetect=off
option wouldn't be understood, or would be pointless.
Hardly a expert on GRES configuration, so just spitballing here...
Prentice
On 2/23/21 3:39 PM, Prentice Bisbal wrote:
How many nodes are we talking about here? What if you gave each node
it's own gres.conf file, where all of them said
AutoDetect=nvml
Except the one you want to exclude, which would have this in gres.conf :
NodeName=a1-10 AutoDetect=off Name=gpu File=/dev/nvidia0
It seems to me like Autodetect and Autodetect=off are exclusive in the
same gres.conf file, but maybe my suggestion would work. If you have a
small number of GPU nodes, or use a configuration management tool like
Ansible, Chef, or Puppet, it might be worth a shot.
Prentice
On 2/19/21 11:31 AM, Paul Brunk wrote:
Hi all:
(I hope plague and weather are being visibly less than maximally cruel
to you all.)
In short, I was trying to exempt a node from NVML Autodetect, and
apparently introduced a syntax error in gres.conf. This is not an
urgent matter for us now, but I'm curious what went wrong. Thanks for
lending any eyes to this!
More info:
Slurm 20.02.6, CentOS 7.
We've historically had only this in our gres.conf:
AutoDetect=nvml
Each of our GPU nodes has e.g. 'Gres=gpu:V100:1' as part of its
NodeName entry (GPU models vary across them).
I wanted to exempt one GPU node from the autodetect (was curious about
the presence or absence of the GPU model subtype designation,
e.g. 'V100' vs. 'v100s'), so I changed gres.conf to this (modelled
after 'gres.conf' man page):
AutoDetect=nvml
NodeName=a1-10 AutoDetect=off Name=gpu File=/dev/nvidia0
I restarted slurmctld, then "scontrol reconfigure". Each node got a
fatal error parsing gres.conf, causing RPC failure between slurmctld
and nodes, causing slurmctld to consider the nodes failed.
Here's how it looked to slurmctld:
[2021-02-04T13:36:30.482] backfill: Started JobId=1469772_3(1473148) in batch
on ra3-6
[2021-02-04T15:14:48.642] error: Node ra3-6 appears to have a different
slurm.conf than the slurmctld. This could cause issues with communication and
functionality. Please review both files and make sure they are the same. If
this is expected ignore, and set DebugFlags=NO_CONF_HASH in your slurm.conf.
[2021-02-04T15:25:40.258] agent/is_node_resp: node:ra3-6 RPC:REQUEST_PING :
Communication connection failure
[2021-02-04T15:39:49.046] requeue job JobId=1443912 due to failure of node ra3-6
And to the slurmd's :
[2021-02-04T15:14:50.730] Message aggregation disabled
[2021-02-04T15:14:50.742] error: Parsing error at unrecognized key: AutoDetect
[2021-02-04T15:14:50.742] error: Parse error in file /var/lib/slurmd/conf-cache/gres.conf
line 2: " AutoDetect=off Name=gpu File=/dev/nvidia0"
[2021-02-04T15:14:50.742] fatal: error opening/reading
/var/lib/slurmd/conf-cache/gres.conf
Reverting to the original, one-line gres.conf reverted the cluster to
production state.
--
Prentice Bisbal
Lead Software Engineer
Research Computing
Princeton Plasma Physics Laboratory
http://www.pppl.gov
--
Prentice Bisbal
Lead Software Engineer
Research Computing
Princeton Plasma Physics Laboratory
http://www.pppl.gov
