Presentation | 2023-12-05 Evaluation of conversion overheads for the sparse matrix format appliying indices of the non-zero elements dictionary compression to accelerate SpMV on GPU Shun Murakami, Kazunori Yoneda, Iwamura Takashi, Masahiro Watanabe, Yasushi Inoguchi, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | In recent years, as numerical simulations have become increasingly complex and large-scale. There is a growing demand for fast computation of linear systems with sparse matrices whose almost elements are mostly zeros and whose have more than several million rows. Iterative methods that do not deform the coefficient matrices are often used to solve these equations. And GPUs, which have faster memory bandwidth than CPUs, have been used to accelerate sparse matrix vector products (SpMV), which take up a major part of the computation time. GPUs have been used to accelerate the sparse matrix-vector multiplication (SpMV). In storing a large sparse matrice to the limited device memory of GPUs, the memory-efficient CSR format is well used. In addtion, SELL-C-σ format has been proposed to improve the memory access pattern and enable fast SpMV computation. Therefore, at the 190th HPC conference, we proposed a non-zero element position dictionary compressing sparse matrix format that can compute SpMV faster on GPUs by applying dictionary compression to non-zero element indices and reducing memory accesses. The proposed storage format is up to 19.6% faster than the CSR format. Although these improved formats speed up the time of SpMV, they cause the overhead of format conversion time. In this paper, we evaluate the SpMV speedup, including the overhead of format conversion, about the CSR format, the SELL-C-σ format, and the proposed non-zero element indices dictionary compressed sparse matrix format on CPU and GPU. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Sparse Matrix-Vector multiplication / GPU / Sparse Matrix Format |
Paper # | CPSY2023-31 |
Date of Issue | 2023-11-28 (CPSY) |
Conference Information | |
Committee | CPSY / IPSJ-ARC / IPSJ-HPC |
---|---|
Conference Date | 2023/12/5(2days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | Okinawa Industry Support Center |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | Computer Systems, HPC, etc. |
Chair | Kota Nakajima(Fujitsu Lab.) / 津邑 公暁(名工大) / Takahiro Katagiri(名大) |
Vice Chair | Yasushi Inoguchi(JAIST) / Tomoaki Tsumura(Nagoya Inst. of Tech.) |
Secretary | Yasushi Inoguchi(Univ. of Tsukuba) / Tomoaki Tsumura(Hitachi) / (富士通) / (九大) |
Assistant | Ryuichi Sakamoto(Tokyo Inst. of Tech.) / Takumi Honda(Fujitsu) |
Paper Information | |
Registration To | Technical Committee on Computer Systems / Special Interest Group on System Architecture / Special Interest Group on High Performance Computing |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | Evaluation of conversion overheads for the sparse matrix format appliying indices of the non-zero elements dictionary compression to accelerate SpMV on GPU |
Sub Title (in English) | |
Keyword(1) | Sparse Matrix-Vector multiplication |
Keyword(2) | GPU |
Keyword(3) | Sparse Matrix Format |
1st Author's Name | Shun Murakami |
1st Author's Affiliation | Japan Advanced Institute of Science and Technolog(JAIST) |
2nd Author's Name | Kazunori Yoneda |
2nd Author's Affiliation | Section Solutions Div., , Solutions Development Unit, Fujitsu Japan Limited(Fujitsu Japan) |
3rd Author's Name | Iwamura Takashi |
3rd Author's Affiliation | Section Solutions Div., , Solutions Development Unit, Fujitsu Japan Limited(Fujitsu Japan) |
4th Author's Name | Masahiro Watanabe |
4th Author's Affiliation | Section Solutions Div., , Solutions Development Unit, Fujitsu Japan Limited(Fujitsu Japan) |
5th Author's Name | Yasushi Inoguchi |
5th Author's Affiliation | Japan Advanced Institute of Science and Technolog(JAIST) |
Date | 2023-12-05 |
Paper # | CPSY2023-31 |
Volume (vol) | vol.123 |
Number (no) | CPSY-293 |
Page | pp.pp.25-30(CPSY), |
#Pages | 6 |
Date of Issue | 2023-11-28 (CPSY) |