Federated Learning with Individualized Differential Privacy

Sync files to the cluster

rsync -avu . oe152msue@login01.sc.uni-leipzig.de:~/flidp

Add --delete as option to delete files that do not exist in the source directory.

Resource Access on Cluster

To interactively allocate resources use salloc, e.g.

salloc -N 1 -p clara --gres=gpu:rtx2080ti:1 --mem=32G --cpus-per-task=8 --time=6:00:00

Possible GPU choices are gpu:rtx2080ti:1 and gpu:v100:1

Then run any command with srun. To free the resources type exit or cancel the job via scancel.

To check the status of your submitted jobs run

squeue -u $USER

Pull docker image

To log into ghcr run

apptainer remote login --username <USER> docker://ghcr.io

To pull an image run

singularity pull docker://<ghcr-image>

Checks during the execution

To check the GPU utilization during a run, you can run this command on the SLURM cluster (here for a running job with id 9189302)

srun --jobid 9189302 --pty nvidia-smi

Runs

MNIST -> ~/logs/flidp-9189302/stdout.out