cluster_101

Some minimal examples on how to submit job in a SLURM-based or CONDOR-based computing clusters.

CONDOR

To find out which Machines have GPUs installed you can run:

condor_status -constraint 'PartitionableSlot && TotalGpus > 0' -af:h Machine TotalGPUs TotalCpus CUDADeviceName TotalMemory CUDAGlobalMemoryMb CUDACapability

Interactive job

condor_submit_bid 25 -i
condor_submit_bid 25 -i -append request_cpus=2 -append request_memory=4096
condor_submit_bid 25 -i -append request_cpus=4 -append request_gpus=8 -append request_memory=4096

Launch a job via submission fle

condor_submit_bid 25 hello_condor.sub

Sweeps

condor_submit_bid 25 condor_sweep.sh

A good example Another example

SLURM

Get infos about cluster nodes:

sinfo -o "%20N %10c %10m %20f %20G %10P"
sinfo -o "%20N %10c %10m %20f %20G %10P" | sort | uniq -c

Interactive job

srun --partition=gpu --gres=gpu:1 --time=00:15:00 --cpus-per-task=4 --pty bash

Launch a job via submission fle

sbatch hello_slurm.sh

Sweeps

sbatch slurm_sweep.sh

A good example

TODO:

• add a small training example
• add SLURM ✅
• lighter conda env
• conda sourcing
• better path def/expansion