Five researchers from Argonne National Laboratory have won the Ewing “Rusty” Lusk Best Paper award at EuroMPI/USA 2025, held at the University of North Carolina in October. EuroMPI is an event for discussion of new developments and applications of the Message Passing Interface (MPI) standard for high performance parallel computing.
The award-winning paper, “Implementing True MPI Sessions and Evaluating MPI Initialization Scalability,” was co-authored by Hui Zhou, Ken Raffenetti, Mike Wilkins, Yanfei Guo and Rajeev Thakur.
“We are deeply honored to receive this award,” said Zhou. Zhou is the lead author and a principal software development specialist in the MCS division. “Rusty would be proud to see us continuing to improve MPICH and MPI for today’s scientific community.”
The paper focused on a new feature in the MPI 4 standard called the Sessions model, which enables an MPI process to locally initialize and construct communicators on demand. Sessions removed the overhead associated with the global MPI communicator comm_world — especially helpful for large-scale applications.
“MPICH has supported MPI Sessions since 2021,” said Raffenetti, a principal specialist, research software engineering in the MCS division. “But it didn’t fully achieve the goal of decoupling MPI completely from the world process model.”
And so, the Argonne team undertook a major refactoring of MPICH to enable it to support what they call a “true” Sessions model.
The paper focused on two main areas: building the new model and measuring how well it works. For the implementation they developed a more efficient approach that uses hierarchical bootstrapping — a powerful technique for analyzing data at multiple levels. For example, the new approach exploits group-level address exchange and shared memory.
“Our results show that the implementation is important,” Zhou said. “With dense communication the benefits are limited. But with sparse communication the new model is faster and uses less memory.”
They also found that the new model greatly improves MPICH’s dynamic process management. The “true” Sessions model, which does not depend on the MPI comm_world communicator, requires an arbitrary group of MPI processes to be able to bootstrap into a fully optimized communicator without information from comm_world. This is essentially the same requirement of MPI dynamic processes. The work of supporting “true” Sessions in MPICH leads to a future performance boost for applications that use MPI dynamic processes. Dynamic processes are commonly used in scientific workflow applications to provide flexibility, resilience and resource malleability.
The researchers noted that developing a feature such as “true” Sessions is a complex, cyclical process — from design to testing to real-world use (see Figure 1).
“But done right, with excellent feedback at each stage, it can give scientists a powerful new tool for solving complex problems,” Zhou said.
For the abstract and slides from the presentation at EuroMPI/USA 2025, see https://eurompi.org/assets/slides/Zhou-True-Sessions.pdf
The award is named after Ewing “Rusty” Lusk (1943–2022), a founder of the Message Passing Interface Forum. Hechaired MPI-2 and co-started the MPICH project to create the first implementation of the MPI standard. He also served as director of the Mathematics and Computer Science (MCS) division at Argonne from 2005 to 2011 and was an Argonne Distinguished Fellow Emeritus.
