Round about 150 new concepts for MPI
Computer racks of SuperMUC-NG
Innovations for supercomputers at software and hardware level also change programming languages and standards. The fourth major release of the Message Passing Interface (MPI) was published in the summer. In about 150 new functions, it contains fundamentally new concepts for the exchange of messages between computing nodes and thus for increasing the performance of high-performance computers (HPC). The first practical experiences with the MPI 4 standard were a topic during EuroMPI21, but progress continues: users and researchers already discussed the next performance level and how MPI can be prepared for the coming generation of exascale systems during the conference in September. With an international team, Prof. Dr. Martin Schulz from Technical University Munich (TUM), Director of Leibniz Supercomputing Centre, and Dr. Martin Ruefenacht, staff member of the (LRZ), put together the EuroMPI21 programme. They report on the innovations and the future of MPI.
How did EuroMPI21 go? Dr Martin Ruefenacht: Very well overall - more than 120 people attended EuroMPI, and more than 45 people attended the following MPI Forum on Standardisation. These are record numbers, which were certainly helped by the virtual format. Prof. Dr Martin Schulz: EuroMPI was planned to be shorter this year - one conference day spread over two afternoons. The content was more focused on MPI standardisation. This was obviously very well received. The second day was dominated by the topics of remote memory access and the MPI sessions, both of which play a major role in the future of MPI. This resulted in interesting and lively discussions about future concepts.
The most important topic was certainly the major version 4 of MPI ... Schulz: MPI 4 is an important milestone and introduces about 150 fundamentally new concepts and methods, such as persistent collectives, large-counts and partitioned communication or new modules for fault tolerance as well as MPI sessions or the new interface for performance analyses. The possibilities they bring for users and new codes were of course intensively discussed. But MPI 4 has only been in use since June 2021 - it was probably too early for a comprehensive assessment of advantages and disadvantages, so ideas for version 5 came to the fore.
What surprised you most as organisers? Why? Schulz: The submissions had led to the formation of two focal points in terms of content: Remote Memory Access with MPI and extensions to the new concept of MPI sessions. We didn't expect this, but the contributions already showed the directions in which MPI could develop. Very exciting. Ruefenacht: Actually, the EuroMPI is a comparatively small conference, but the number of participants this year was surprisingly high. Probably more people were able to listen in the afternoon and online than just those discussing standardisation during the MPI Forum.
What's next for MPI - are there already ideas for MPI 5 and when can we expect them? Ruefenacht: Of course there are already ideas to develop MPI4 further. But it will take several years until the next major version. The MPI Forum is currently concentrating on MPI 4.1, which is something of a clean-up operation. We will improve the language, clarify and correct ambiguities and perhaps add a few new features. In preparation for MPI 5, I am participating in a programming language working group. MPI 5 could be specified for more languages than C or FORTRAN. At the moment, the focus is on C++ because more and more HPC applications are based on it and therefore the demand for a good interface to MPI is growing. Schulz: In addition, we see new approaches for the topics of remote memory access, MPI sessions and for areas such as partitioned communication. These will have a positive effect on programming that addresses GPU and CPU at the same time and also bring new interfaces for performance enhancement and for debugging tools. So MPI is going from strength to strength.
Prof. Dr. Martin Schulz, TUM und LRZ; Dr. Martion Ruefenacht, LRZ