Seems if they are really that short, it would be better to have a single
job run through them all, or 10 jobs run through 2000 each kind of thing.
Such short jobs take more time for setup/teardown than the job itself,
making this approach inefficient. The amount of resources used just to
schedule them in that fashion outweighs the resources needed by far.
Brian Andrus
On 8/28/2020 3:30 AM, navin srivastava wrote:
Hi Team,
facing one issue. several users submitting 20000 job in a single batch
job which is very short jobs( says 1-2 sec). so while submitting more
job slurmctld become unresponsive and started giving message
ending job 6e508a88155d9bec40d752c8331d7ae8 to queue.
sbatch: error: Batch job submission failed: Unable to contact slurm
controller (connect failure)
Sending job 6e51ed0e322c87802b0f3a2f23a7967f to queue.
sbatch: error: Batch job submission failed: Unable to contact slurm
controller (connect failure)
Sending job 6e638939f90cd59e60c23b8450af9839 to queue.
sbatch: error: Batch job submission failed: Unable to contact slurm
controller (connect failure)
Sending job 6e6acf36bc7e1394a92155a95feb1c92 to queue.
sbatch: error: Batch job submission failed: Unable to contact slurm
controller (connect failure)
Sending job 6e6c646a29f0ad4e9df35001c367a9f5 to queue.
sbatch: error: Batch job submission failed: Unable to contact slurm
controller (connect failure)
Sending job 6ebcecb4c27d88f0f48d402e2b079c52 to queue.
even that time the load of cpu started consuming more than 100% of
slurmctld process.
I found that the node is not able to acknowledge immediately to
server. it is moving from comp to idle.
so in my thought delay a scheduling cycle will help here. any idea how
it can be done.
so is there any other solution available for such issues.
Regards
Navin.
