The online LUMI Roadshow attracted 67 participants from different disciplines in academia as well as public administration and private enterprises. The overall large interest resulted in an interesting Q/A session about LUMI’s features, functionality and accessibility options both from potential future users and support staff who will be assisting Swedish LUMI users. Below is an aggregated summary of this Q/A session.
Regarding the features of the LUMI supercomputer, Dr. Pekka Manninen, Director – LUMI Leadership Computing Facility, stated that the GPU performance will reach 550 PFLOPS, while the CPU partition has a peak performance of about 8 PFLOPS. The CPU-only jobs can request anything between one to 100000 cores. Concerning the module system, the LUMI team is looking at how the Cray PE works with Lmod, with the fallback option of using Cray’s own module system.
With respect to supported frameworks and programming models, Pekka also noted that Tensorflow and PyTorch are supported ML frameworks (see https://www.amd.com/en/graphics/servers-solutions-rocm-ml) and Python and Julia are naturally parts of the computing environment and supported programming models, while LUMI-D will be the platform for Spark and similar elastic computing software environments. Also the container cloud service can be used, although it is not meant for large-scale computing. There are also plans to make build systems, such as EasyBuild and spack, available in order to facilitate application deployment on the machine. Researchers can use software preinstalled on the system or build their own software – support for software builds will be available from the user support team.
Licences for commercial software are also included in LUMI’s total budget according to Pekka, i.e. pre-paid for the end-users, though not unlimited, so some prioritisation needs to be done by the user support team. It will also be possible for users to use their own software licenses.
The number of projects expected on LUMI is heavily dependent on the country and the allocator. Tens of simultaneous Tier-0 allocations are expected plus some hundreds of smaller projects. The largest jobs (up to half or even the whole machine) will need human-level intervention to drain jobs from the machine and arrangements will need to be made with systems administrators.
To handle various maintenance of LUMI, the system vendor will bring a sysadmin team to be based in Kajaani. CSC is also hosting other Finnish supercomputers there, and synergies will be found between the administration of those and LUMI.
LUMI and Sweden
Researchers in Sweden will be able to interact with LUMI through the following channels
- ENCCS – Help with proposal application
- ENCCS – Help with medium to extensive porting to GPUs
- LUST/SNIC AE – Help adjusting compilation parameters, fixing smaller portability issues, optimisation etc
The Swedish part of LUMI will be a separate call, open only for Swedish researchers. It will be announced via the ENCCS and SNIC mailing lists,web pages and other social channels. The EU part of LUMI and other systems will be announced via a new EuroHPC JU portal similar to the current PRACE portal and will also be distributed via ENCCS and SNIC mailing lists.
How LUMI resources are positioned in the country’s research infrastructure is a national-level decision (SNIC in case of Sweden). In principle the system and policies are designed so that use cases are not blocked out. The criteria for being granted a project via the national part of LUMI will include showing the specific need for LUMI as well as knowledge on how to use resources on this scale efficiently.
The EU procedures are still in discussion, however they are not expected to be much different from current PRACE procedures.
Both ENCCS and SNIC have already announced courses relevant to LUMI, for example the workshop on OpenMP for GPU offloading which will be given by ENCCS in June (follow the ENCCS events page!) and a CSC course on HIP programming and porting.
The first line of support for LUMI will go through the LUMI User Support Team (LUST). There will be just one support portal, but system-level aspects (accounts, logins, storage allocations etc) will be forwarded to the operational staff directly.
If you missed the LUMI Roadshow, you can get the summary and slides at https://enccs.se/events/2021/01/lumi-roadshow/. SAVE the date for the next one on 8th of April.
Below you can subscribe to our newsletter for quarterly updates on our work and future events.
Learn more about LUMI on https://www.lumi-supercomputer.eu.