On 31/01/18 20:28, Miguel Gutiérrez Páez wrote:
If I unlimit memory resources (by commenting the last line in
custom.conf file), the same sbatch works properly. A scontrol show
job of a failed job shows that the job was launched in a compute
node, where there is no any restriction about memory (or other)
resource. So, the login node is the only node I limit resources. Why
is failing the sbatch if the compute nodes have no any restriction
but the login one?
By default Slurm propagates resource limits from the node you are
submitting from. Check the PropagateResourceLimits section of the
slurm.conf manual page.
Short version, add this to slurm.conf:
PropagateResourceLimits NONE
Hope that helps!
All the best,
Chris
--
Chris Samuel : http://www.csamuel.org/ : Melbourne, VIC
