Hi Ole.
I'm using the packages from Debian stable (slurm 20.11.4, hwloc 2.4.1).
And I checked: hwloc is installed on all the nodes. Quite obvious since
it's a dep for slurmd:
https://packages.debian.org/bullseye/slurmd
Being a dep, i "suspect" slurmd is built with hwloc support.
Diego
Il 07/11/2021 20:22, Ole Holm Nielsen ha scritto:
Hi Diego,
Are you sure that the Slurm software installed on all compute nodes was
actually built on a system which had the hwloc packages installed? They
should also be installed on the compute nodes. The prerequisite
packages are listed here:
https://wiki.fysik.dtu.dk/niflheim/Slurm_installation#install-prerequisites
/Ole
On 05-11-2021 15:38, Diego Zuccato wrote:
They aren't using modules so it must be something system-wide :(
But not all jobs are impacted. And it seems it's a bit random (doesn't
happen always).
I'm out of ideas, currently :(
Il 05/11/2021 13:10, Ole Holm Nielsen ha scritto:
On 11/5/21 12:47, Diego Zuccato wrote:
Some users are reporting this error:
slurmstepd-str957-mtx-01: error: hwloc_get_obj_below_by_type()
failing, task/affinity plugin may be required to address bug fixed
in HWLOC version 1.11.5
slurmstepd-str957-mtx-01: error: task[0] unable to set taskset '0x0'
I checked on that node and hwloc is newer:
diego.zuccato@str957-mtx-01:~$ hwloc-info --version
hwloc-info 2.4.1
How can Slurm detect such an old HWLOC version?
Maybe the user loads a software module which also loads an old hwloc
module? The user should do "module list" in the job to verify this.
--
Diego Zuccato
DIFA - Dip. di Fisica e Astronomia
Servizi Informatici
Alma Mater Studiorum - Università di Bologna
V.le Berti-Pichat 6/2 - 40127 Bologna - Italy
tel.: +39 051 20 95786
