miniHPC: SMALL BUT MODERN HPC

Funding: University of Basel

Duration: 15.12.2016-Present

Project Summary

miniHPC is a small homogeneous CPU cluster of the HPC Group (High Performance Computing) at the Department of Mathematics and Computer Science at the University of Basel. It serves the purpose of a platform for teaching parallel programming to students, so that they learn to achieve high performance computations.

miniHPC consists of 21 identical dual-socket compute nodes based on Intel Xeon E5-2640 v4 (Broadwell) processors, together with a login node and a storage node, all connected by a dual-rail 100 Gbit/s Intel Omni-Path fabric. Because every compute node has the same hardware, miniHPC is well suited to teaching MPI and OpenMP and to experiments where uniform, reproducible hardware matters. The cluster has a theoretical peak performance of about 16.1 double-precision TFLOP/s. The GPU and large-memory nodes that were previously part of miniHPC are now part of HPCaTS.

Additional Notes

• Intel Xeon E5-2640 v4 is a 10-core Broadwell processor with a 2.4 GHz base clock (up to 3.4 GHz turbo) and AVX2/FMA vector units. Each node has two sockets, one NUMA domain per socket.
• Peak performance: 16 double-precision FLOP/cycle per core × 2.4 GHz × 20 cores = 768 GFLOP/s per node, or about 16.1 TFLOP/s for the 21 compute nodes.
• Dual-rail Intel Omni-Path (two Omni-Path HFI 100 Series adapters per node, 100 Gbit/s each) carries the compute traffic between nodes. A separate Gigabit Ethernet network handles user access and administration.
• ~55 TB of shared storage served over NFS from a dedicated storage node and mounted on every node.
• Slurm workload manager with a single compute partition.
• Ubuntu 24.04 LTS, with software provided through Lmod environment modules, including OpenMPI 4.1.6 (with and without UCX).