English

OpSparse: a Highly Optimized Framework for Sparse General Matrix Multiplication on GPUs

Distributed, Parallel, and Cluster Computing 2022-06-16 v1

Abstract

Sparse general matrix multiplication (SpGEMM) is an important and expensive computation primitive in many real-world applications. Due to SpGEMM's inherent irregularity and the vast diversity of its input matrices, developing high-performance SpGEMM implementation on modern processors such as GPUs is challenging. The state-of-the-art SpGEMM libraries (i.e., nsparsensparse and spECKspECK) adopt several algorithms to tackle the challenges of global load balance, local load balance, and allocation of the result matrix. While these libraries focus on the high-level algorithm design for SpGEMM, they neglect several low-level architecture-specific optimizations, which causes inefficient implementations in their libraries. In this paper, we classify their inefficient implementations into seven categories. Based on our observations, we propose a highly optimized SpGEMM library called OpSparseOpSparse. The optimizations in OpSparseOpSparse include 1) optimizing the binning method by improving the utilization of the shared memory, 2) optimizing the hashing method by reducing the access to the hash table, 3) improving the trade-off between hash collision rate and hardware utilization in the hashing method by setting appropriate binning ranges, 4) reducing the overheads of global memory utilization by minimizing the global memory usage of the metadata, and 5) improving the execution parallelism by overlapping global memory allocation with kernel execution. Performance evaluations with 26 commonly used matrices on an Nvidia Tesla V100 GPU show that OpSparseOpSparse achieves up to 27.8×27.8\times, 1.81×1.81\times, and 2.04×2.04\times performance speedup over three state-of-the-art libraries: cuSPARSEcuSPARSE, nsparsensparse, and spECKspECK, respectively.

Keywords

Cite

@article{arxiv.2206.07244,
  title  = {OpSparse: a Highly Optimized Framework for Sparse General Matrix Multiplication on GPUs},
  author = {Zhaoyang Du and Yijin Guan and Tianchan Guan and Dimin Niu and Linyong Huang and Hongzhong Zheng and Yuan Xie},
  journal= {arXiv preprint arXiv:2206.07244},
  year   = {2022}
}

Comments

This paper has been submitted to the IEEE Access since May 7, 2022, and is currently under review by IEEE Access. 20 pages, 11 fgures, 5 tables

R2 v1 2026-06-24T11:51:42.520Z