Workshops Archive

End-to-end GPU-native communication in GROMACS simulations with PME task specialization


Workshop: The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM)

Authors: Andrey Alekseenko (KTH Royal Institute of Technology), Mahesh Doijade (NVIDIA), and Szilárd Páll (KTH Royal Institute of Technology)

Abstract: Strong scaling of molecular dynamics simulations is increasingly limited by communication latency as per-step execution time becomes shorter. In task-specialized GROMACS runs, particle-particle (PP) tasks exchange coordinates and forces with particle-mesh Ewald (PME) tasks every time step through CPU-mediated MPI communication, exposing host synchronization on the critical path of a step that is otherwise entirely GPU-resident. We present an NVSHMEM-based GPU-native redesign of these pipelined producer-consumer communication phases. Our design combines stream-triggered and kernel-initiated communication, fuses PME force communication with receiver signaling, and is tuned for NVLink and RDMA interconnects to reduce control-path overhead and increase overlap, improving end-to-end performance by up to 15%. This work extends GPU-native communication beyond halo exchange toward end-to-end GPU-resident GROMACS execution, while highlighting the challenges that symmetric allocation without team-based support poses for task-specialized applications.


Back to The 9th Annual Parallel Applications Workshop, Alternatives to MPI+X (PAW-ATM) Archive Listing

Back to Full Workshop Archive Listing