Be prepared for Euro-HPC resource LUMI
Here we present facts about Euro-HPC resource LUMI (Finish for snow) to be built in Finish Kajaani and available to Swedish researchers from 2021. Images and text presented here are taken from the ENCCS event https://enccs.se/events/2021/01/lumi-roadshow/
LUMI will be an HPC Cray EX supercomputer with peak perfomance of 550 petaflop/s that can be compared to world-fastest Fugaku in Japan with 513 petaflop/s or second fastest Summit in USA with 200 petaflop/s
LUMI capabilities
- Extreme computing capacity based on LUMI-G and LUMI-C partitions
- LUMI queue policies will support jobs from single CPU core of a GPU to 50% of the nodes (even 100% with special arrangements)
- Jobs can combine resources from both sides within a workflow, even with the same executable
- Interactive use (visualization, data analysis, pre/post processing,..) on LUMI-D
- Broad stack of pre-installed scientific software, databases and datasets, both commercial and community
- Sharing datasets over LUMI-O service
- Running microservices on LUMI-K
- Expore quantum computing with LUMI-Q
Enhanced user experience
- Besides command-line interface, Jupyter Notebooks, Rstudio will be integrated
- Large software budget enable rich stack of pre-installed software
- Dataset as service i.e. curated large reference datasets available and maintained
- Support for handling sensitive data (GDPR subjected, IP-closed, etc)
How to prepare for LUMI
- Thinking projects and use cases for LUMI
- Cases for Tier-o grand challenges
- Combining simulation and AI methods within the same workflow
- There is a vast pool of GPU-enabled community codes
- See if your favorite software suite already has bee enabled - if not consider moving to a competing packate that is
- Perhaps only part of the application need to be GPU-enabled, rest running on the CPU nodes
- Modernizing applications and GPU-enabling them "even if it works, fix it"
LUMI programming environment
- ROCm (Radeon Open Compute)
- Standard set of accelerated scientific libraries (BALS, FFT etc)
- Standard machine learning framework and libraries (Tensorflow, PyTorch etc)
- Compilers for the GPUs
- Cray Programming Environment (CPE) stack
- Cray Compiling Environment, LibSci libraries, CrayPAT, Reveal, debuggers,...
- CPE Deep Learning plugin
Prepare applications and workflows for LUMI
- Remember the possibility of combining CPU and GPU nodes within one job - perhaps only part of application needs to be GPU-enabled
- Consider writing your application on top of modern frameworks and libraries
- Kokkos, Alpacka etc, or domain-specific frameworks
- Convert CUDA codes to HOP, OpenACC codes to OpenMP5
- HIPify tools can automatize the effort
- LUMI phase1 will compe with a code porting platform (MI100 GPUs)
- HIP porting can be done already now on Nvidia GPU platforms
LUMI preparations
ENCCS is dedicated to help Swedish researchers preparing for LUMI and available on:
