Skip to content

Koios

Koios is the previous-generation CEICO cluster. It has 27 compute nodes, each with two 16-core Intel Xeon Skylake CPUs (Xeon Gold 6130) and 384 GB of RAM. All nodes run Rocky Linux 9.8.

GPU node temporarily unavailable

The GPU node with four NVIDIA Tesla P100 cards is currently dedicated to a project and is not available through Slurm. It will return to general use later. Until then, use the A100 nodes on Phoebe for GPU work.

Accessing Koios

Log in through the front-end node koios1.fzu.cz (see SSH). Koios shares accounts and SSH keys with Phoebe, so the same username and key work on both. Its home directories are separate from Phoebe's.

Slurm partitions

Jobs on Koios are managed by Slurm. Jobs go to the cpu partition unless you ask for another one with --partition.

Partition Nodes Per node Time limit Use
cpu (default) n[1-9,11-27] 32 cores (64 threads), 384 GB 9 days 1 h batch jobs
cpu_int n[11-12] 32 cores (64 threads), 384 GB 9 days 1 h interactive work
preempt n8 32 cores (64 threads), 384 GB 9 days 1 h jobs that may be preempted
small_int s1 (virtual machine) 8 cores, 23 GB 7 days 7 h jobs that need only a small node

Limits change from time to time; sinfo on the login node shows the current values.

Hardware

Node type Amount Hostnames Processors GPUs Cores (logical CPUs) Main memory NVMe
compute 27 n[1-27] 2× Xeon Gold 6130 - 32 (64) 384 GB 1× 2 TB
GPU 1 gpu2 2× Xeon Gold 6130 4× Tesla P100 SXM2 16 GB 32 (virtual) 293 GB -

The GPU node is a virtual machine. It uses 32 of the 64 logical CPUs of its physical server, and the four GPUs are passed through to it directly. It is currently dedicated to a project (see the note at the top of this page).

The nodes are connected by 100 Gb/s InfiniBand EDR (Mellanox MT4115 ConnectX-4 cards), a low-latency network for communication between nodes. More detail: Koios infrastructure.

Storage and software

Koios has its own home directories, software tree (/cvmfs/c9.phoebe.lan) and shared scratch (/mnt/shared-scratch); the project space /mnt/proj is the same as on Phoebe. Each job gets a private /tmp on the node's NVMe disk. See storage and software.