- edited description
SLURM_NODELIST formatted funny: node[572-578]
Hi,
In my case, the SLURM_NODELIST environment variable looks like node[572-578]. Line 84 in amp/utilities.py already deals with node[572, 578] so the problem is trivial. scontrol show hostname $SLURM_NODELIST seems another clean solution.
Also, amp.utilities.assign_cores doesn't seem to take precedence over the automated guesses, which prevents a user from fixing the problem by providing a custom cores dictionary.
By the way, ntasks-per-node might be empty as well - I usually use ntasks.
Thank you.
Best regards, Xiang
Comments (5)
-
reporter -
repo owner Thanks -- we essentially got this working in our own queuing system, but rely on others to make sure it works for them.
Perhaps you can provide a patch? The scontrol show hostname sounds like a clean solution, as you say.
-
reporter Thanks. I am no expert in Python, Git or SLURM, but here is a patch on
utilities.py
.scontrol show hostname
will hopefully convertnode[01-02]
tonode01,node02
ornode01\nnode02
, which is then parsed usingre.split
. -
reporter - attached 0001-initial-branching.patch
-
reporter - changed status to resolved
- Log in to comment