.png){kind=icon}
返回
Optimizing sparse matrix partitioning in a heterogeneous CPU-GPU system for high-performance
.png){kind=icon}
PRE
AI
DOI：10.1007/s00607-025-01456-5.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Using graphics cards to enhance application performance while reducing energy consumption has become increasingly prevalent. In the context of CPU-GPU systems, where the GPU operates as an accelerator, the efficient distribution of workloads between these processors becomes a pivotal factor that directly affects the performance of application execution. This challenge assumes even greater significance and complexity within applications that involve sparse matrix processing. This complexity arises from the differing architectures of these processors and the unique characteristics of these applications, which fall under the category of irregular applications. This article introduces a strategy for partitioning sparse matrix on a CPU-GPU heterogeneous platform, with the primary objective of achieving better performance. We present a framework for distributing irregular workloads to the CPU and regular workloads to the GPU. We evaluate our proposed algorithms using the Sparse Matrix-Vector Multiplication (SpMV) kernel on a system equipped with an NVIDIA GeForce RTX 1070 GPU. Evaluation of partitioning on real-world sparse matrices demonstrates that the proposed algorithms provide significant performance improvement. However, it is essential to emphasize that none of the proposed algorithms can be definitively considered the optimal solution. Our experimental results, obtained from real-world sparse matrices, illustrate that the performance of these algorithms is contingent upon the distribution of non-zero elements and the degree of sparsity. Therefore, selecting the most suitable algorithm should be based on a comprehensive evaluation of these factors.
Keyword:
Partitioning
Irregularity
SpMV
GPU
