Accelerating SpMV for Scale-Free Graphs with Optimized Bins

Sparse matrix-vector multiplication ($SpMV$) is a fundamental operation in numerous scientific applications, particularly in the context of graph analytics. As graph-based computations become increasingly complex, there is a growing demand for the development of more efficient Sp MV. In this paper, we present a novel approach called Binn to enhance SpMV performance for scale-free graphs on modern multicore processors. Binn incorporates three key optimizations to accelerate SpMV. Firstly, it employs an adaptive cache blocking strategy, which partitions the adjacency matrix of a graph into 2D blocks of varying sizes. This promotes balanced workloads and cache efficiency. Secondly, Binn reorders the nonzero elements of the adjacency matrix, enabling regularized access patterns within each block. Lastly, Binn identifies and eliminates redundant message passing during the execution of SpMV, resulting in reduced memory costs. Through these optimizations, Binn aims to accelerate$SpMV$by facilitating efficient data movement across the memory-cache hierarchy and achieving workload balance among threads. Experimental evaluation on diverse graph datasets demonstrates the effectiveness of Binn, outperforming state-of-the-art Sp MV implementations and graph systems such as Intel's MKL by$3.78\times$and Galios by$1.47\times$.