A Sparse Matrix Optimization Method for Graph Neural Networks Training.

Graph neural networks (GNN) have shown great application potential in scientific research applications, biomedicine, and other fields, which exhibit superior feature representation capabilities for graph data with non-Euclidean structures. These capabilities are enabled efficiently by sparse matrix-matrix multiplication (SPMM) and sparse matrix-vector multiplication (SPMV) that operate on sparse matrix representations of graph structures. However, SpMM has the characteristics of high memory occupation and irregular memory access, which leads to low storage and computational efficiency. To address the above issues, this paper proposes a sparse matrix optimization method, including a sparse matrix format and a performance model. The format, namely BMCOO, divides the sparse matrix into multiple blocks and adopts the bitmap to compress the position information of non-zero elements in each block. This paper further designs an SpMV algorithm in BMCOO format on GPU. In addition, a multi-channel SpMV performance model is constructed to predict the execution time of SpMV by combining the sparse matrix scale and system architecture parameters. Then the performance model fine-tunes the graph partitioning of the GNN training process. Experiments on the SuiteSparse and the Open Graph Benchmark datasets verify the effectiveness and superiority of the proposed method.