Workshop: PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems
Authors: Antoine Kempf, Ansar Calloo, and Cédric Chevalier (French Alternative Energies and Atomic Energy Commission, Université Paris-Saclay); Romain Le Tellier (French Alternative Energies and Atomic Energy Commission); and Marc Perache (French Alternative Energies and Atomic Energy Commission, Université Paris-Saclay)
Abstract: Discontinuous Galerkin (DG) discretizations of diffusion-type operators, including neutron transport, produce sparse systems that bottleneck simulation throughput. Standard incomplete LU (ILU) preconditioners exhibit O(h−1 ) iteration growth and prohibitive fill-in. This work presents a Smoothed Aggregation Algebraic Multigrid (SA-AMG) preconditioner for 3D DG systems, implemented via Trilinos/MueLu and Kokkos, demonstrating O(1) mesh-independent convergence on the Takeda Benchmark Model 1 up to 7.1e6 unknowns, yielding a 158× speedup over ILU. A strong scaling and roofline analysis across three manycore topologies — Intel Sapphire Rapids, AMD Genoa, and NVIDIA Grace — shows all solver phases memory-bound. On both dual-socket nodes, scalability regresses or halts at the socket boundary due to inter-NUMA traffic in the fine-grid solve, while the single-socket Grace platform sustains its efficiency and achieves the lowest time-to-solution despite fewer cores. These results identify memory topology, rather than core count, as the dominant scalability factor for heterogeneous reactor transport calculations.
Back to PMBS26: The 17th International Workshop on Performance Modeling, Benchmarking, and Simulation of High Performance Computer Systems Archive Listing
Back to Full Workshop Archive Listing