Workshops Archive

Load-Balanced Communication Overlap for Lattice H-Matrix–Vector Multiplication on GPU Clusters


Workshop: Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026)

Authors: Naoki Momotake and Tetsuya Hoshino (Nagoya University), Akihiro Ida (Japan Agency for Marine-Earth Science and Technology), So Ozawa and Ryo Ando (The University of Tokyo), Masatoshi Kawai (Tohoku University), and Takahiro Katagiri (Nagoya University)

Abstract: Lattice hierarchical H-matrices provide a structured low-rank representation for parallel boundary element method solvers, but the concentration of dense and poorly compressible submatrices near the matrix diagonal causes load imbalance under conventional process-grid mappings. We develop a lattice-H-matrix-specific design for load-balanced communication overlap on GPU clusters. The design combines a one-dimensional column-cyclic task assignment with a coarse-grained diagonal/off-diagonal decomposition. The assignment distributes the expensive diagonal-block workload across processes, while the decomposition allows the collective reduction of off-diagonal contributions to proceed concurrently with diagonal-block GPU kernels. An MPI+OpenACC implementation was evaluated on NVIDIA GH200 nodes using the HBI earthquake simulation application. The proposed design distributes the dominant diagonal-region workload more evenly and achieves up to a 3.08x speedup over the conventional two-dimensional process-grid implementation.


Back to Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026) Archive Listing

Back to Full Workshop Archive Listing