1 of 9

Introduction to LUMI

Slides by LUST

Several slides kindly contributed by Orian Louant, LUMI User Support Team

2 of 9

LUMI and the LUMI consortium

  • HPE-Cray EX supercomputer featuring AMD CPUs and GPUs
  • Funded by the EuroHPC JU (50%) and a consortium of 10 countries
  • Resources of LUMI allocated per investment
    • Norway’s share about 2 %
  • The EuroHPC share will be allocated via a peer review process (similar to PRACE) and available for all European researchers
  • The remaining share will be allocated by the consortium members by local consideration

3 of 9

LUMI right now

  • 1536 compute nodes with 2x AMD EPYC 7763 (Milan)
  • 128 cores per node, 196 608 cores total
  • 256 GiB of memory per node
  • Some nodes with 512 GiB and 1 TiB
  • Cray Slingshot interconnect
  • 4 x 20 PB Lustre filesystems (80 PB total) (LUMI-P)
  • 1 Lustre flash filesystem (7 PB) (LUMI-F)

Phase 1: CPU Partition (LUMI-C)

Pilot phase October 20th – December 31st

Ca 40 projects from the 10 consortium countries

Regular operation since January 1st, 2022

​

4 of 9

LUMI in the future

  • Cloud partition based on Kubernetes (LUMI-K)
  • Object storage based on Ceph (LUMI-O)
  • Data analytics partition (high-memory and NVIDIA RTX8000) (LUMI-D)

Phase 2: GPU Partition (LUMI-G)

Also available in the future

  • Available summer 2022
  • Next generation AMD Instinct GPUs (2560 nodes with 4 x MI250X)
  • 550 Pflops peak performance
  • Will be the fastest supercomputer in Europe

5 of 9

The LUMI User Support Team

  • Helpdesk
  • User documentation
  • Computing environment
  • User training
  • Porting and optimization consultancy

A distributed team in charge of:

Emmanuel Ory

Christian Schou Oxvig

Jørn Dietze

6 of 9

Getting access to LUMI & EuroHPC

​

7 of 9

How to prepare for LUMI-C

LUMI-C: Betzy has somewhat similar architecture / user interface

  • Differences
    • Betzy: AMD Rome (zen2) vs LUMI-C: AMD Milan (zen3)
    • Betzy: InfiniBand vs LUMI-C: Slingshot network
    • Betzy: Large software stack on system vs LUMI-C: very limited stack
    • LUMI-C users should be very comfortable with installing software themselves
  • EasyBuild, Slurm, Lustre should be known for Betzy users

​

​

8 of 9

How to prepare for LUMI-G

LUMI-G: a “beast” unlike anything else

  • May require significant porting & tuning effort before applications even run
  • Not many hardware resources to start porting working currently available (only on previous GPU generation MI100 vs MI250X)
  • Some good documentation on hardware available at https://docs.olcf.ornl.gov/systems/crusher_quick_start_guide.html#crusher-quick-start-guide (no guarantee that LUMI-G will be identical in all aspects)
  • If you are familiar with CUDA, OpenMP (Offloading) or OpenACC or other programming frameworks (SYCL, Kokkos, Raja), you might be “well” prepared.
  • In any case, contacting the NRIS GPU team is strongly recommended!
  • Open call for help with porting software to LUMI-G and AMD GPUs

​

9 of 9

How to … LUMI?

Contacts

Resources

​