SHIRO: Near-Optimal Communication Strategies for Distributed Sparse Matrix Multiplication

Chen Zhuang 1,2 , Lingqi Zhang 2 , Benjamin Brock 3 , Du Wu 1,2 , Peng Chen 2 , Toshio Endo 1 , Satoshi Matsuoka 1,2 , Mohamed Wahib 2 · 2025

Read SHIRO: Near-Optimal Communication Strategies for Distributed Sparse Matrix Multiplication by Chen Zhuang 1,2 , Lingqi Zhang 2 , Benjamin Brock 3 , Du Wu 1,2 , Peng Chen 2 , Toshio Endo 1 , Satoshi Matsuoka 1,2 , Mohamed Wahib 2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-20178.txt.

Description

Chen Zhuang leads an eight-author team cutting wasted chatter from distributed sparse matrix work. Sparse matrix-matrix multiplication backs HPC and deep learning alike. On clusters, its bottleneck is communication. Two flaws feed that cost. Transfers ignore the matrix's sparsity pattern. Layouts waste the network's tiers, pushing data over slow links between nodes for no need.

SHIRO answers with a fine-grained, sparsity-aware communication plan. It sends only what the pattern requires. The plan is mapped onto two-tier GPU network layouts, to keep traffic on fast links. Tests on real data scaled to one hundred twenty-eight GPUs. At that scale, speedups over four top baselines ran from about nine-fold to over two-hundred-fold, in geometric mean. The work was accepted at ICS 2026.

Catalog record

CREATORChen Zhuang 1,2 , Lingqi Zhang 2 , Benjamin Brock 3 , Du Wu 1,2 , Peng Chen 2 , Toshio Endo 1 , Satoshi Matsuoka 1,2 , Mohamed Wahib 2
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-20178

Browse the full library · About the Tower of Babel