CIIRC Computational Cluster Description
The CIIRC computational cluster serves CIIRC researches, students, and their collaborators for intensive batch computations. The CIIRC cluster also offers acceleration on graphics cards. We have been extending and improving the CIIRC cluster.
The CIIRC cluster is XSEDE Compatible Basic Cluster based on OpenHPC project. The CIIRC cluster runs on CentOS 7 operating system. It uses Slurm workload manager and a job scheduler. Cluster management and orchestration is done by Warewulf toolkit.
The core of the cluster consists of five NVIDIA DGX-1 nodes suplemented by other gpu and computational nodes. See table below for further details:
|# of nodes||23|
|RAM Total||7 TB|
|# of V100-32-MaxQ GPU||40|
|# of GTX 1080 Ti; 11 GB GPU||18|
|Theoretical CPU performance||28 TFLOPS|
|Theoretical GPU performance SP||750 TFLOPS|
Nodes are interconnected by 10 Gbit Ethernet. The 100 Gbit EDR IB (InfiniBand) is available. All nodes contain a local ssd scratch disk. The whole cluster is connected to the 600 TB Isilon NAS storage. „compute*“ means a default partition for computations without gpu acceleration. „GRES: means generic resources in Slurm.
Present Slurm configuration is fairly standard (simplified):
Future extension and plans:
- BeeGFS distributed scratch space
- The cluster throughput improvement by better QOS settings
Building and continuous extending of the CIIRC computational cluster needs substantial financial resources. The money comes mainly from national and European Commission projects, e.g. the Czech Government investment which created CIIRC, Josef Urban’s ERC Consolidator project AI4REASON and his project AI & Reasoning; Josef Šivic’s project IMPACT, Robert Babuška’s project R4I, etc.
Responsible: Jan Kreps; Poslední změna 2020-11-19