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,
and two GPU nodes, gpu3 and gpu4, each with two NVIDIA Tesla P100
cards. The login node, the GPU nodes and most compute nodes run Rocky Linux 10; the eight compute
nodes in the legacy partition run Rocky Linux 9.8 until they are reinstalled.
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; use it for batch and interactive CPU work.
| Partition | Nodes | Per node | Time limit | Use |
|---|---|---|---|---|
cpu (default) |
n[1-19] |
32 cores (64 threads), 384 GB | 9 days 1 h | batch and interactive jobs |
legacy |
n[20-27] |
32 cores (64 threads), 384 GB | 9 days 1 h | Rocky Linux 9.8, until these nodes are reinstalled |
small_int |
s1 (virtual machine) |
8 cores, 23 GB | 7 days 7 h | jobs that need only a small node (currently out of service) |
gpu |
gpu[3-4] (virtual machines) |
2 Tesla P100, 32 cores, 125 GB | 9 days 1 h | GPU jobs |
The partitions cpu_int, rocky10 and rocky10_cpu were removed in November 2026 (use cpu),
and preempt is closed.
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 | 2 | gpu[3-4] |
1× Xeon Gold 6130 (virtual) | 2× Tesla P100 SXM2 16 GB | 32 (virtual) | 125 GB | - |
The GPU nodes are two virtual machines on one server with four P100 cards. Each uses one of the server's two CPUs (32 logical CPUs) and the memory attached to it, and gets two GPUs passed through directly.
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 and shared scratch (/mnt/shared-scratch); the project
space /mnt/proj is the same as on Phoebe. On the compute nodes each job gets a private /tmp
on the node's NVMe disk; on the GPU nodes /tmp is a local disk shared by the jobs on that node.
See storage and software.
All software modules on Koios come from the EESSI software stack, on the
login node and on every compute and GPU node. The former Koios module tree /cvmfs/c9.phoebe.lan
has been retired: rebuild programs that used its modules with an EESSI toolchain.