SC21 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Charm++ and AMPI: Adaptive and Asynchronous Parallel Programming

Authors: Sam White (University of Illinois, Charmworks Inc), Laxmikant Kale (University of Illinois, Charmworks Inc), Abhinav Bhatele (University of Maryland), Michael Norman (San Diego Supercomputer Center (SDSC)), Jozsef Bakosi (Los Alamos National Laboratory), Esteban Meneses (Costa Rica National High Technology Center, Costa Rica Institute of Technology), Kavitha Chandrasekar (University of Illinois)

Abstract: This BoF is community gathering about parallel programming using Charm++, Adaptive MPI, the many applications built on them and associated tools, covering recent advances in Charm++ and the experiences of application developers with Charm++ and AMPI. There will also be a discussion on the future directions of Charm++, opportunities to learn more and to form collaborations. Charm++ is a production-grade many-tasking programming framework and runtime system for modern HPC systems. It offers high productivity and performance portability through features such as multicore and accelerator support, dynamic load balancing, fault tolerance, latency hiding, interoperability with MPI and OpenMP, and online job-resizing.

Long Description: The goal of the Charm++/AMPI BoF is to provide a forum to discuss recent developments in Charm++ and Adaptive MPI, for application teams to report results and experiences, to shape future directions of the programming system, and to encourage education and collaboration. Charm++ embodies a leading-edge, production implementation of the adaptive, asynchronous, multitasking programming model that is becoming increasingly important and widespread. It is used in over a dozen large-scale, production HPC applications, including NAMD, which won the 2002 Gordon Bell prize and 2012 Sidney Fernbach award. The combined usage of Charm++ applications on current US supercomputers is second only to MPI codes. Several Charm++ applications are part of 'early-science' projects for next-generation supercomputers under development. AMPI is an MPI implementation on top of Charm++ that provides the familiar API of MPI with the dynamic runtime features of Charm++. Charm++/AMPI BoFs have been held at SC every year since 2012. It has consistently attracted a diverse audience ranging from people who had never used Charm++ or AMPI to expert developers with multiple publications on them. Attendees have included developers of the Charm++ system, researchers from a range of universities, government laboratory staff, and commercial application developers and users. It is an important community-building event for those interested in alternatives to MPI programming. The SC19 BoF was attended by about 60 people. The SC18 BoF and SC17 BoF were each attended by about 70 people. The SC16 BoF was attended by about 80 people, with more coming into the room and leaving when they saw that it was standing-room-only. The SC15 BoF overfilled the allotted 65-seat room, with 10 members of the development team forced to step out to make room for other attendees. At least 15 more prospective attendees were turned away at the doors due to lack of space. The SC14 BoF had approximately 60 attendees.


Back to Birds of a Feather Archive Listing