Skip to content

The cluster at a glance

This page lists the values of the SC cluster that this guide uses. They were checked on 20 October 2026 with sinfo and scontrol on the cluster and against the SC documentation, and they can change. sinfo -s on the login node shows the partitions as they are today.

Machines

Short name Host What for
sc-login login01.sc.uni-leipzig.de starting jobs and looking at them, never computing
sc-export export01.sc.uni-leipzig.de moving files to and from the cluster
— portal.sc.uni-leipzig.de your account, your SSH key, support requests
— lab.sc.uni-leipzig.de Jupyter in the browser
— rstudio.sc.uni-leipzig.de RStudio in the browser
— www.sc.uni-leipzig.de the SC documentation

The names sc-login and sc-export come from the SSH settings that the setup guide writes in step 5.

These are the fingerprints of the servers' keys, the same as on the SC page SSH host keys:

Host Type Fingerprint
login01 ED25519 SHA256:3saqgn+bVpG3PyG7mtwM2nPQY0NfK9auNCxLCWb85wQ
login01 ECDSA SHA256:IOSVuQ/ip4FiXKZicXAYChm0KiibigU10KHzTREbhrs
login01 RSA SHA256:02tNZqOPx+eqt7kIZbJaw2Vz6vaK7sj8iBbwm5yNyj8
export01 ED25519 SHA256:5lwMCnUQYG50VHNRKULsTdcF1tHGm+FZ8duSEJTppNA
export01 ECDSA SHA256:xYK3VypMR/nMYJFLi6yvrHtKearKMSxzlznxYacZe68
export01 RSA SHA256:GqM5sPXxQ0RNoTkqwVO+m11ATEYaK4pmKsBBYX2qY7A

Partitions

The table shows the computers of each partition and their sizes as sinfo reports them. The SC page Hardware describes them in more detail.

Partition Machines Cores per machine Memory per machine GPUs per machine Memory per core, at most
cpu-xeonv3 (default) 80, t001 to t080 16 or more 125 GB — 12.5 GB
cpu-zen3 32, paul01 to paul32 128 500 GB — 93.75 GB
cpu-zen1 10, polaris01 to polaris10 32 248 GB — 15.6 GB
cpu-bigmem 1, sirius01 128 5.9 TB — 93.75 GB
gpu-a30 12, paula01 to paula12 128 981 GB 8 × A30, 24 GB each 15.3 GB
gpu-l40s 4, firehorse01 to firehorse04 144 376 GB 4 × L40S 31.25 GB
gpu-v100 6, clara 32 500 GB 4 × V100, 32 GB each 31.25 GB
gpu-rtx2080ti 18, clara 32 500 GB 8 × RTX 2080 Ti, 11 GB each 31.25 GB

The code-alongs use cpu-xeonv3 and gpu-a30. The SC documentation still uses older names for the same computers, such as paul, paula, clara, polaris, sirius and t3000. sinfo lists both kinds of name.

GPUs

A job script books a GPU with #SBATCH --gpus=TYPE:COUNT, for example --gpus=a30:1. The types are a30, l40s, v100 and rtx2080ti.

PyTorch from PyPI brings CUDA 13, which needs GPU driver 580 or newer and does not support the V100, as a notice of PyTorch explains. The A30 and L40S have driver 580.178.04, so the GPU code-alongs book an A30.

Limits

Slurm treats each run as a job of its own. So these limits, and the wait as PENDING, apply to every run: to each line of runs.txt, or to the single run of a job without it.

Limit Value
time of a run at most 2 days (--time=2-00:00:00); 1 hour if the job script sets no --time
memory of a run 1 GB per booked core if the job script sets no --mem; the most per core is in the table of partitions
runs of a job at most 99,999 (MaxArraySize=100000)
end at the time limit Slurm sends the program the signal TERM and ends it 30 seconds later (KillWait)

A job that asks for more than its partition allows is refused by sbatch at once and never waits (EnforcePartLimits=ANY).

Each core comes with a share of memory, as the table of partitions shows. A job that asks for more memory than its cores come with also gets more cores, as the SC page Slurm explains.

Longer jobs run in the partitions whose names end in -long, such as cpu-zen3-long with 10 days or cpu-zen1-long with 42 days. They are open only to groups that SC lets in, and sinfo -a -s lists them. Without access, sbatch stops with User's group not permitted to use this partition. Ask SC support for access.

Modules

This guide uses two modules:

Module Used by
R/4.5.2-gfbf-2025b first-jobs/hello-r.slurm, the R code-alongs
Python/3.12.3-GCCcore-13.3.0 first-jobs/hello-python.slurm, the code-along on long runs

module spider R on the login node lists all versions of R, and the SC software catalogue lists all software. The Python code-alongs with packages install their own Python with uv instead.