Container does not start on a small set on cluster

[lidavid88]

I have discovered a problem on running container on a cluster.

I am using a nvidia pytorch container created with enroot in the following submit script:

#!/usr/bin/env bash

#SBATCH --time=03:00:00
#SBATCH --gres=gpu:4
#SBATCH --nodes=8
#SBATCH --gpus-per-node=4
#SBATCH --ntasks-per-node=4
#SBATCH --mem=501600mb

ml purge

run_num='0'

SRUN_PARAMS=(
  --mpi="pmi2"
  --gpus-per-task=1
  --gpu-bind="closest"
  --label
  --container-name=fcn
  --container-mounts=/etc/slurm/task_prolog.hk:/etc/slurm/task_prolog.hk,/scratch:/scratch,/hkfs/work/workspace/scratch/usr1234,/tmp,/usr/bin/srun:/usr/bin/srun
  --container-mount-home
  --container-writable
  --no-container-entrypoint
)

srun "${SRUN_PARAMS[@]}" bash -c "
  echo $run_num
"

On most nodes srun is executed and I get 0 printed to the log.

But on the other nodes I get 2 types of errors:

1.

22: slurmstepd: error: pyxis: container start failed with error code: 1
22: slurmstepd: error: pyxis: printing enroot log file:
22: slurmstepd: error: pyxis:     /etc/enroot/hooks.d/10-shadow.sh: line 70: 3474706 Broken pipe             yes 2> /dev/null
22: slurmstepd: error: pyxis:          3474707 Segmentation fault      (core dumped) | pwck -R "${ENROOT_ROOTFS}" "${pwddb#${ENROOT_ROOTFS}}" /etc/shadow > /dev/null 2>&1
22: slurmstepd: error: pyxis:     nvidia-container-cli: ldcache error: process /usr/sbin/ldconfig failed with error code: 1
22: slurmstepd: error: pyxis:     [ERROR] /etc/enroot/hooks.d/98-nvidia.sh exited with return code 1
22: slurmstepd: error: pyxis: couldn't start container
22: slurmstepd: error: spank: required plugin spank_pyxis.so: task_init() failed with rc=-1
22: slurmstepd: error: Failed to invoke spank plugin stack

2.

21: slurmstepd: error: pyxis: couldn't start container
21: slurmstepd: error: spank: required plugin spank_pyxis.so: task_init() failed with rc=-1
21: slurmstepd: error: Failed to invoke spank plugin stack

This error does not appear, if I only use up to 4 nodes.

With 8 nodes the job works, if I am lucky. But most of the time I get errors on some nodes.

My guess is that the inter node communication is having troubles with pyxis.

Can someone help me with that?

Regards

[flx42]

It the issue happens 0% of the time on some nodes and 100% of the time on some nodes, I suggest you start investigating the differences between the good nodes and the bad nodes:

• Is it the same distro, Linux version, NVIDIA driver version?
• Is it the same enroot version? Perhaps try to reinstall enroot on the bad nodes.
• Check dmesg and the slurmd log on the bad nodes for any clue.

[lidavid88]

They seem to have the same versions and drivers.