Skip to content

Koios: two Tesla P100 GPU nodes in the gpu partition

Since 16 October 2026 Koios has two GPU nodes again, gpu3 and gpu4, each with two NVIDIA Tesla P100 16 GB cards, in the new Slurm partition gpu. They replace the former GPU virtual machine that held all four cards and was dedicated to a project.

How to use them

Submit from koios1.fzu.cz, for example:

srun --partition=gpu --gres=gpu:p100:1 --cpus-per-task=16 --time=01:00:00 --pty bash

GPU jobs has batch and interactive examples and the software on the nodes. In short:

  • The P100 (Pascal, compute capability 6.0) is not supported by CUDA 13. The nodes have CUDA 12.9 from the NVIDIA HPC SDK; build with nvcc -arch=sm_60.
  • The nodes run Rocky Linux 10, and the Koios software tree /cvmfs/c9.phoebe.lan is not available there. Programs built on the Koios compute nodes may need to be rebuilt.
  • Home directories, /mnt/shared-scratch and /mnt/proj are the same as on the rest of Koios.

How it is set up

The four P100 cards sit in one server. Each node is a virtual machine on it with one of the server's two CPUs (32 logical CPUs), the memory attached to that CPU (125 GB) and two GPUs passed through directly; the two GPUs of a node are connected by one NVLink.

On slurmserver2.koios, slurm.conf got the two node lines with Gres=gpu:p100:2, the partition gpu (Nodes=gpu[3-4], open to all, MaxTime=9-01:02:03), the nodes in SuspendExcNodes, and AccountingStorageTRES=gres/gpu, so sacct now records GPU use on Koios. The GPU type comes from NVML autodetection (gres.conf AutoDetect=nvml). The nodes run Slurm 25.11, the controller's release series.