I have access to an HTC. I want to run ntasks-per-node=32 parallel instances of the same python script on 1 node. Here is the slurm submit file at the moment:

I have the ff sbatch script:

I have the ff sbatch script:

I cannot use GPU sharding, even though everything seems to have been configured according to the instructions:

I want to find out when my jobs are going to start. According to the docs it should be possible with squeue --start, however the start times seem to be N/A until the job starts, and it also is just a date. I would like to get an estimate, according to the current state of the queue, in how many minutes/hours/days is my job going to be executed using SLURM.

I want to find out when my jobs are going to start. According to the docs it should be possible with squeue --start, however the start times seem to be N/A until the job starts, and it also is just a date. I would like to get an estimate, according to the current state of the queue, in how many minutes/hours/days is my job going to be executed using SLURM.

When I checked with the code “lscpu”, it shows

I have two slurm partitions (lhpc and lgpu) with a shared node (n16-90). I have configured one partition with higher priority. I want that if one job uses the shared node through the lgpu partition and there is already one job in the lhpc partition, the latter is suspended and the former allocates the shared node.

When the MIG is disabled, you can specify a graphics card device by running /dev/nvidia[0-7], but after the MIG device is enabled, the MIG device cannot be found.

Currently, I have multiple GPU nodes and pool them through slurm. Enroot.conf adopts the default configuration. At this time, the image pulled by enroot can only be cached on the same node. When running a task on another node, you need to Re-pulling the image results in a waste of time.