Frequently Asked Questions¶

Download the account request form, complete sections 1 and 2, sign it, and send it to your institution's HR representative (or department secretary for universities). They will verify your affiliation, complete section 3, and forward it to accounts-hpc@dipc.org. See HPC Accounts for the full procedure.

Submit a new account request form following the same procedure as the initial request, indicating the new expected end date at your institution. See Account Renewal for details.

Only if you collaborate with a research group that has its own computational resources at the DIPC. In that case, fill out the external collaborator form and send it to accounts-hpc@dipc.org. See Access for external collaborators.

Contact us at accounts-hpc@dipc.org and we will reset your credentials.

Open a terminal and run ssh username@atlas-fdr.sw.ehu.es (replace username with your assigned username). See How to Connect for OS-specific instructions and configuration tips.

Generate a key pair on your local machine with ssh-keygen -t ed25519, then upload your public key to account-setup.sw.ehu.es via SSH. This is mandatory: accounts without a registered key cannot access the clusters. See Account Setup for step-by-step instructions.

This means SSH cannot find or use your private key. Verify that your key exists in ~/.ssh/ and that file permissions are correct (chmod 600 ~/.ssh/id_ed25519). If using MobaXterm or another client, make sure the session is configured to use your private key. See Troubleshooting.

This usually happens after a maintenance intervention that changes the server's host key. Remove the old key with ssh-keygen -R atlas-fdr.sw.ehu.es and try connecting again.

Yes. Create a ~/.ssh/config file on your local machine to define shortcuts. See SSH config for examples.

Run squeue -j <jobid> -o "%.18i %.30j %.10R" to see the reason. Common causes: Priority (fair-share; your priority increases over time), Resources (not enough free nodes right now), ReqNodeNotAvail (a requested node is in DRAIN or otherwise unavailable), or QOSMaxWallDurationPerJobLimit (your --time exceeds what the QoS allows). If the reason is Resources and the cluster is not busy, your request may exceed what any single node can provide. See Troubleshooting for the full list of reasons, and Job priority and scheduling for how the Priority value is computed.

The reason field is the last explanation SLURM recorded when it evaluated your job, not a real-time status. It refreshes only when the scheduler looks at that job again, so it can lag behind the actual cluster state. Your job stays in the queue and is considered every scheduling cycle regardless of what the field reads. See The reason in squeue looks stuck or outdated for how to inspect node states and what to do.

Make sure you submitted your job from the /scratch filesystem, not from your home directory. SLURM writes output files to the working directory from which the job was submitted, so if that directory is not writable or is on a different filesystem, the output may not appear where you expect.

/dipc is your home directory; on Atlas FDR and Atlas EDR it is backed up daily and is meant for important files, scripts, and results you want to keep. /scratch is high-speed temporary storage for running jobs. /scratch is not backed up and should only be used for data that is actively being processed. Move important results to /dipc after your job completes.

Yes, but only of the home directories of Atlas EDR and Atlas FDR. We perform daily backups of /dipc. Data deleted within the past two months can typically be recovered. Contact support-hpc@dipc.org with the file paths and approximate deletion date. We do not perform backup of the home directories located on Hyperion cluster.

Unfortunately, no. We do not back up the scratch filesystems. The /scratch directory is intended for temporary, active-use data only. Always copy important results to a permanent storage location after your jobs complete.

On Hyperion, use getquota-home, getquota-scratch, and getquota-data. On Atlas EDR and Atlas FDR, use getquota (shows all quotas at once). Both space and file count (inode) usage are displayed. See Storage and quotas for details.

Find what is using the most space with du -sh /scratch/<username>/* | sort -rh | head -20, then clean up common offenders: conda clean --all, apptainer cache clean, pip cache purge, or old Nextflow work directories (nextflow clean -f). If you need more space temporarily, contact us at support-hpc@dipc.org. See Storage and quotas for the full checklist.

/lscratch is a local disk on each compute node. It is faster than /scratch because I/O does not go through the network, but it is only visible from that node and is not persistent. Use it when your job creates or reads many small files (e.g., temporary databases, image datasets). The typical pattern is to copy input to /lscratch at the start, work there, and copy results back to /scratch when done. See Local scratch for I/O-intensive jobs for an example script.

No. Atlas and Hyperion have separate home and scratch directories. To move data between them, transfer through your local machine (download from one, upload to the other) or set up SSH keys between the clusters for direct transfer. See Data Transfer for details.

Our clusters have no predefined group directories, but you can set one up yourself on /dipc, /scratch, or /data (Hyperion). Use standard Unix group permissions when everyone shares a Unix group, or ACLs for per-user access or mixed groups. See Sharing a directory with collaborators for the step-by-step recipes.

For most cases, use rsync. It handles large transfers efficiently, resumes interrupted transfers, and only sends what has changed. For quick one-off copies, scp works well. For interactive browsing, use sftp or a GUI client like WinSCP. See Data Transfer for examples and a comparison table.

Try enabling compression (rsync -avz or scp -C). If possible, schedule large transfers during off-peak hours. Also check that your local network connection is stable. Transfers over WiFi can be significantly slower than wired connections.

We use the Lmod module system. Run module avail to see all available software, and module load <name> to load a specific package. Use module spider <name> to search for a specific module and see its dependencies.

Contact us at support-hpc@dipc.org with the software name, version, and a link to the project page. We will evaluate whether it can be installed system-wide. Alternatively, you can install software in your home directory or use containers (Apptainer/Singularity).

Send an email to support-hpc@dipc.org. Please include your username, the cluster you are using, and a clear description of the issue (including error messages and job IDs if applicable).

Contact us at support-hpc@dipc.org, we are happy to help.