John Biddescombe
Visualization/Computational Scientist, CSCS
Integrating asynchronous MPI messaging with tasking runtimes requires careful handling of request polling and completion dispatching. Using C++26 senders, we implemented MPI messaging in the pika library as execution graph nodes. This enables flexible scheduling and execution, improving performance. We present insights on design choices affecting polling, scheduling, and completion handling and present results from Alps under various configurations.