Benchmarking and developing NAMD With Charm++ and Reconverse
Date:
NAMD relies on Charm++ for its distributed capabilities such as dynamic load balancing and GPU-offload mode. In this talk, we review the impending change of the Charm++ communication and scheduling layer from old Converse to Reconverse. We explore the (very few) code changes required for NAMD to support Charm++ with Reconverse, the immediate benefits of switching to Reconverse, and our future plans for using new Reconverse features such as device-to-device communication and adaptive queue scheduling to improve NAMD’s scaling performance.