The Koios cluster, a previous-generation system, is composed of 28 compute nodes. Each node is equipped with two 16-core Intel Xeon Skylake CPUs. Moreover, a dedicated node is allocated for GPU-accelerated computing, hosting four high-performance Nvidia Tesla P100 GPUs.
Koios is available using koios1.fzu.cz front-end SSH node.
Koios shares the authentication backend with Phoebe, meaning that the same username and SSH key are sufficient for logging in.
Jobs at Koios are managed by the Slurm workload manager, which implements multiple fair-use and priority-queuing algorithms for effective system utilization.
(detailed list of Koios hardware is here)
| Node type | amount | Hostnames | Processors | GPUs | Number of cores (logical CPUs) | Main memory | nvme |
|---|---|---|---|---|---|---|---|
| compute | 27 | n[1-27] |
2x Xeon Gold 6130 | - | 32(64) | 384 GB | 1x 2TB |
| GPU | 1 | n28 |
2x Xeon Gold 6130 | 4x Tesla P100 | 32(64) GPU:4x3584 | 384 GB | 1x 2TB |
The nodes are equipped with MT4115 ConnectX-4 1x100Gb Infiniband EDR cards, ensuring a low-latency network for efficient inter-node communication.
User homedirs are located at Beegfsmore_info cluster filesystem shared with Phoebe.
Each node features a ~2TB Intel NVME, mounted as /tmp and accessible as $TMPDIR. It's important to note that this storage is non-persistent and undergoes periodic cleaning during node image updates.
In the /mnt/archive directory, slower yet deduplicated and compressed (using VDO more_info ) storage is mounted, serving as an archive for infrequently accessed files
Application software is provided using CVMFS (CernVM File System more_info) in /cvmfs. Software is built and organized using Easybuild frameworkmore_info utilizing Lmod software modules.