Remote execution

Quantum chemistry, molecular docking, and MD/GROMACS jobs can run on a remote x86-64 Linux host while the GUI remains on your computer. SilicoLab uses the operating system’s ssh and scp clients and deploys a checksum-verified, version-matched worker on first use, caching it under ~/.silicolab/workers so later deployments — including to new hosts — work without network access.

Set up SSH

Open Settings > Engines > Remote Hosts, add the host’s address, SSH user, port, and work directory, then select Set up passwordless login. SilicoLab creates a dedicated key under ~/.silicolab/keys and keeps strict host-key verification enabled. Run the displayed authorization command once on the remote host, then select Verify.

The work directory defaults to ~/.silicolab. It must be visible and writable on every node that may execute a job. On a cluster this normally means a shared home, project, or scratch filesystem. Select Test scheduler after configuring Slurm; the test submits a real short job and verifies worker visibility from the allocated node.

Job environment commands run inside the allocated job before the worker. Use them for commands such as module load gromacs or CUDA setup. Scheduler setup commands run on the login node before sbatch, squeue, scontrol, and scancel; use them only when Slurm commands are not already on the non-interactive SSH PATH.

Direct SSH

Choose Direct SSH for a dedicated workstation or bare compute node. The worker runs as a detached process group. CPU requests limit the worker thread pool; memory and walltime remain advisory because no scheduler enforces them.

Slurm

Choose Slurm and configure the fields required by the cluster:

• Partition selects a queue, such as debug or gpu.
• Account identifies the allocation charged for the job.
• QOS selects a site-defined quality-of-service policy.
• Reservation and Constraint are optional advanced selectors.
• Default CPU, memory, and walltime values apply when a task does not override them.

Select Detect Slurm to verify sbatch, squeue, scancel, and scontrol. sacct is optional; when it is unavailable, SilicoLab uses scontrol for terminal history. Refresh cluster loads partition, GPU-type, and feature suggestions. These values are hints and do not reserve currently idle hardware.

GRES is the default GPU dialect. A task can request:

• No GPU;
• Any available GPU with a count; or
• Specific type with a scheduler GPU type such as a100 and a count.

SilicoLab translates these requests to --gres=gpu[:type]:count. Select the --gpus dialect only when the cluster administrator recommends it. A site whose GPU request looks like neither can use Custom, a template argument with {count} and an optional {type} placeholder — for example --gres=accel:{type}:{count}. Slurm, not SilicoLab, chooses physical device indices and exposes the allocation to the job.

Run and monitor jobs

Choose the host under Run on in a task panel, then set CPU, memory, walltime, and GPU intent. The task monitor shows queued, running, completing, cancelling, and terminal states. For queued Slurm jobs it also shows scheduler reasons such as Priority, Resources, InvalidAccount, or InvalidQOS.

Use Refresh Remote to retrieve the latest state and appended console output. Closing SilicoLab does not stop a remote job. Reopen the project and refresh to continue monitoring or retrieve its result. The scheduler and remote directory captured at submission remain authoritative even if the host settings are edited later.

Select Cancel for a queued or running job. Slurm cancellation remains in Cancelling until the scheduler confirms CANCELLED; repeated requests are safe. Remote scratch can be removed only after a terminal state is confirmed.

Login-node CPU and GPU utilization is not shown as cluster utilization for a Slurm target. Allocation state and pending reason belong in the task monitor.

Troubleshooting

• Invalid account, QOS, or partition: copy the exact values supplied by the cluster administrator and refresh the job to see Slurm’s pending reason.
• Worker is not visible on the compute node: move the work root to a shared filesystem and run Test scheduler again.
• No terminal history: sacct may be disabled. SilicoLab automatically uses scontrol; controller retention still determines how long old jobs remain observable.
• Typed GPU remains pending: confirm the type spelling under Refresh cluster and check that the selected partition contains that GRES type.
• GROMACS is missing: add the appropriate module or environment command to Job environment commands, then use Detect GROMACS.
• ORCA is not configured: enter the ORCA executable path for that remote host under Compute targets. ORCA is not auto-detected.
• Worker download fails (offline, proxy, or GitHub rate limit): the version-matched worker is fetched from GitHub the first time you deploy a given SilicoLab version, then cached under ~/.silicolab/workers for reuse. To use remote compute offline, run one remote job while online after each update to warm the cache. A cached worker is reused without any network request; only the first fetch for a new version needs GitHub.