Best Practice Guide LUMI
Jussi Heikonen, George Markomanolis, Cristian-Vasile Achim, Ezhilmathi Krishnasamy, Abdulrahman A. Azab, Pedro Ojeda-May, Michele Martone, Marcin Krotkiewski, Hicham Agueny, M. G. Barrios Sazo, Ole W. Saastad · Zenodo (CERN European Organization for Nuclear Research) · 2023
The pre-exascale system LUMI is hosted by CSC IT Center for Science and is located at Kajaani, Finland. While LUMI is an acronym (Large Unified Modern Infrastructure) it is also the Finish word for snow. At the time of writing the LUMI system ranks as number 3 on the Top500 list. LUMI is well described in detail by the LUMI consortium. A short summary is presented below. The system was delivered by Hewlett Packard Enterprise (HPE) and is an HPE Cray EX supercomputer. The committed Linpack (HPL) performance of the GPU partition of LUMI is 375 Petaflops/s coming from AMD InstinctTM MI250X GPUs. The system has various partitions of which LUMI-C (x86 partition) and LUMI-G (accelerator partition) are the most important. Other partitions, include storage and other services see. As LUMI is powered by GPUs many codes may need to be adapted to take advantage of the accelerators. As the GPUs are from AMD the software tools are somewhat different from NVIDIA based systems. These and related issues are discussed in the LUMI documentation.