The first pilot phase of the LUMI supercomputer launched on 20 October. In this pilot phase, two pilot projects from each member country in the LUMI consortium participate.
The pilots' task is to test the CPU system before it is opened to ordinary users. From Norway, projects from the Rosseland Center and the NorESM community are participating. This pilot phase will last until the middle of December.
After almost 3 weeks of access to LUMI, the pilot users are now sharing their first impression of the machine.
Experiences from NorESM
The NorESM community has so far tested the containerized version of its software on the machine: — Using singularity makes everything very easy to port on any platform, says Jean Iaquinta, who is one of the pilot users from the NorESM community. He continues: - Overall, we are satisfied with the machine's performance. It is what we expected.
Puhuri, a NeIC-funded project, handles access control and resource management for LUMI. Puhuri's goal is to contribute to seamless access to LUMI and other resources. When asked about the login process, we get the following comment from Anne Claire Foullioux, also a NorESM pilot user: - Login and access to the machine have been very smooth - a lot easier than PRACE systems we have tried. This is great.
The next step in pilot testing for the NorESM community is to complete benchmarking and compare the results with software compiled directly on the machine.
Experiences from Rosseland Center
When it comes to Rosseland’s pilot project, the name of their software is Bifrost. Bifrost is used to simulate solar physics, and in this case, the software is compiled directly on LUMI with the Cray compiler. Mikolaj Szydlarski is one of the Rosseland pilot users and he has shared his experiences with compilation and performance with us:
— It was straightforward to compile Bifrost with the Cray PE compiler. However, performance is a bit disappointing at this point. We almost exclusively work with Intel Fortran compilers, and we know from our experience on Betzy that even for AMD CPUs they can produce fast code. On the other hand, Cray compilers open possibilities for much finer tunning; thus, I am sure we can at least match per core performance with Sigma2's Betzy machine. The code on LUMI is about 30% slower as of now.
Mikolaj continues: — So far, we are pleased, and the whole process has been smooth, and we could quickly start production runs. We also plan to test our new-gen code (Dispatch-Bifrost), which depends on many modern features of MPI, OpenMP and Fortran.” He concludes "The machinery of that code is way more complicated, and we can only hope that tests will go as smooth as with "vanilla" Bifrost.
Extra pilot projects selected
As most of the LUMI pilots are up and running without difficulty, a few selected projects will be added to the LUMI-C pilot from Friday 12 November. We pick these projects from the list of the top projects running on Betzy today.
Mainly GPU-based calculations
Most of the computing capacity of LUMI comes, as previously mentioned, from GPUs, in contrast to more traditional supercomputers where it comes from CPUs. The LUMI machine will be composed of several partitions, including LUMI-C, which consists of CPU nodes and LUMI-G, which consists of GPU nodes.
The pilot testing of LUMI-G is expected to start in March 2022. Here, Rosseland’s Bifrost and GPU code from the Nordic Language Processing Lab at UiO will be pilot projects.
Photo by Fade Creative.