STORM library uses RDMA to scale Monte Carlo transport to 13,440 cores
New open-source library hits 97% weak-scaling efficiency for astrophysical radiation simulations.
Monte Carlo particle transport is critical for high-fidelity astrophysics simulations (supernovae, neutron-star mergers), but traditional two-sided MPI communication creates progress overhead that limits scalability beyond O(10^4) cores. Enter STORM (Scalable Transport via One-sided Remote Memory), an open-source library authored by Maor Mizrachi, Barak Raveh, and Elad Steinberg. STORM replaces MPI's matched send/receive with lock-free, CPU-independent RDMA writes, using per-rank ring buffers to transfer particles without receiver involvement. A two-sided MPI backend offers portability.
In benchmarks, the RDMA backend sustains >97% weak-scaling and >88% strong-scaling efficiency on 13,440 cores (112 cores per network adapter), with speedups of 1.14–1.27x over two-sided MPI. In a Hohlraum IMC benchmark with 4,480 ranks, STORM is 1.41x faster thanks to a 6.1x reduction in MPI progress overhead. By abstracting communication from physics and mesh representations, STORM removes a key barrier to scaling Monte Carlo transport in multiphysics codes, enabling coupled radiation-hydrodynamics with energy- and angle-resolved photon/neutrino transport on evolving meshes.
- STORM uses one-sided RDMA writes instead of two-sided MPI, eliminating CPU involvement for remote rank communication.
- Demonstrates >97% weak-scaling and >88% strong-scaling efficiency on 13,440 cores (112 cores per NIC).
- Reduces MPI progress overhead by 6.1x, achieving up to 1.41x speedup in a Hohlraum simulation benchmark.
Why It Matters
Enables scalable, high-fidelity radiation and neutrino transport for next-generation astrophysical simulations on exascale systems.