Universitätsklinikum Essen ClusterDocsResearch Compute Cluster
Browse documentation

Reference · RCC ClusterDocs

Resources and live discovery

Static node counts, model names, partition names, and memory totals become stale quickly. Discover the schedulable service at the time you submit work.

Related learning: Class 3 explains how to choose resources from measurements rather than static inventory claims.

Scheduler summaries

sinfo
sinfo -o '%P %a %l %D %c %m %G'
scontrol show partition

These commands answer:

  • which partitions are available;
  • their time limits and current state;
  • the CPU and memory shape visible to Slurm;
  • whether GPU resources are advertised.

The current execution model uses cpu_short for jobs up to two hours, interactive for bounded human-in-the-loop sessions, and cpu_nodes for regular CPU batch work including long runners. GPU jobs use the currently advertised GPU partition. Verify live limits rather than assuming that a partition name grants a whole node or unlimited runtime.

Use scontrol show node <allocated-node> only when detailed information about a node assigned to your job is necessary. Do not probe or publish the full physical inventory.

Measure a completed job

sacct -j <jobid> --format=JobID,State,Elapsed,AllocCPUS,ReqMem,MaxRSS,ExitCode

Compare requested resources with actual elapsed time and maximum resident memory. Application-specific profilers remain necessary for CPU scaling, I/O, GPU utilization, and algorithmic bottlenecks.

Choose CPU or GPU

Use a GPU only when the application has a supported GPU implementation and the environment contains compatible user-space libraries. nvidia-smi proves that a GPU and driver are visible; it does not prove that Python, R, CUDA toolkits, or a particular model can use them.

For CPU work, match --cpus-per-task to the application's actual threads. More cores do not accelerate serial code and can lengthen queue time.

Available software

RCC provides a managed baseline including Slurm clients, Miniforge/Conda, Snakemake support, and Apptainer. Verify the executable you will use:

command -v conda mamba snakemake apptainer
conda --version
snakemake --version
apptainer --version

Project software belongs in a versioned Conda environment or reviewed container, not in system directories. If a required runtime or architecture is absent, document the requirement and ask RCC support rather than targeting a physical host by name.

Storage is also a resource

Capacity, throughput, latency, IOPS, metadata operations, and durability are different properties. Shared storage is appropriate for durable input and final results; job-local scratch is appropriate for high-I/O intermediates. Measure the workflow before requesting more CPU, RAM, or GPU resources to compensate for an I/O bottleneck.