Evolve, a collaborative effort between ICL and the University of Houston, expands the capabilities of Open MPI to support the NSF’s critical software-infrastructure missions. Core challenges include: extending the software to scale to 10,000–100,000 processes; ensuring support for accelerators; enabling highly asynchronous execution of communication and I/O operations, and ensuring resilience. Part of the effort involves careful consideration of modifications to the MPI specification to account for the emerging needs of application developers on future extreme-scale systems.
So far, Evolve efforts have involved exploratory research for improving different performance aspects of the Open MPI library. Notably, this has led to an efficiency improvement in multi-threaded programs using MPI in combination with other thread-based programming models (e.g., OpenMP). A novel collective communication framework with event-based programming and data dependencies was investigated. It demonstrated a clear advantage regarding aggregate bandwidth in heterogeneous (shared memory + network) systems. Support for MPI resilience following the User-Level Failure Mitigation (ULFM) fault-tolerance proposal was released based on the latest Open MPI version and will soon be fully integrated into Open MPI.
In Collaboration With
- University of Houston