[slurm-users] slurm only looking in "default" partition during scheduling

[Durai Arasan]

Hi,

We have a cluster with 2 slave nodes. These are the slurm.conf lines describing nodes and partitions:

NodeName=slurm-gpu-1 NodeAddr=192.168.0.200  Procs=16 Gres=gpu:2 State=UNKNOWN
NodeName=slurm-gpu-2 NodeAddr=192.168.0.124  Procs=1 Gres=gpu:0 State=UNKNOWN
PartitionName=gpu Nodes=slurm-gpu-1 Default=YES MaxTime=INFINITE State=UP
PartitionName=compute Nodes=slurm-gpu-2 Default=YES MaxTime=INFINITE State=UP

Running sinfo gives the following:

PARTITION AVAIL  TIMELIMIT  NODES  STATE NODELIST
gpu          up   infinite      1   idle slurm-gpu-1
compute*     up   infinite      1   idle slurm-gpu-2

When I request a gpu job to be run using the following command:

srun --gres=gpu:2 nvidia-smi

I get the error:

srun: error: Unable to allocate resources: Requested node configuration is not available

and in slurmctld.log these are the entries:

[2020-05-12T14:33:47.578] _pick_best_nodes: JobId=55 never runnable in partition compute
[2020-05-12T14:33:47.578] _slurm_rpc_allocate_resources: Requested node configuration is not available 

It seems like slurm is looking only in the partition "compute" and not in the other partitions.
Even if I explicitly specify the gpu node to srun it fails:

srun --nodelist=slurm-gpu-1 nvidia-smi

I get the same error:

srun: error: Unable to allocate resources: Requested node configuration is not available

and in slurmctld.log:

[2020-05-12T14:38:57.242] No nodes satisfy requirements for JobId=56 in partition compute
[2020-05-12T14:38:57.242] _slurm_rpc_allocate_resources: Requested node configuration is not available

It is still looking in partition "compute" even after specifying the node to srun. 

But when I specify a partition, it works:

srun -p gpu nvidia-smi

But I would not like to specify the partition and would like slurm to select nodes based on the options specified in the srun command. Does anyone understand what is wrong in the setup?

Thanks,

Durai

[Michael Robbert]

You have defined both of your partitions with “Default=YES”, but Slurm can have only one default partition. You can see from * on the compute partition in your sinfo output that Slurm selected that one as the default. When you use srun or sbatch it will only look at the default partition unless you use -p

To avoid confusion you should remove the Default=YES from one of the partitions and know that you will need to explicitly ask for the other (non-default) partition whenever you want to use it.

 

Mike

 

From: slurm-users <slurm-use...@lists.schedmd.com> on behalf of Durai Arasan <arasan...@gmail.com>
Reply-To: Slurm User Community List <slurm...@lists.schedmd.com>
Date: Tuesday, May 12, 2020 at 08:49
To: "slurm...@lists.schedmd.com" <slurm...@lists.schedmd.com>
Cc: "benjamin...@uni-tuebingen.de" <benjamin...@uni-tuebingen.de>
Subject: [External] [slurm-users] slurm only looking in "default" partition during scheduling

 

CAUTION: This email originated from outside of the Colorado School of Mines organization. Do not click on links or open attachments unless you recognize the sender and know the content is safe.

[Williams, Jenny Avis]

You cannot have two default partitions.

 

The slurm.conf is picking up the last of the entries flagged as Default; because the compute partition has no partition specified it is being sent to the default partition, thus the first srun is being submitted to the compute partition, and that partition has only a node with no gpu GRES defined, thus the srun is correct to fail.

 

What is the output of the command

sinfo

 

The partition understood to be the default will be indicated with * after the partition name.