面向稀疏矩阵访存特性的Cache划分

稀疏矩阵向量乘是许多科学计算的核心,计算中大量的间接和随机访存成为计算的主要瓶颈。本文通过分析稀疏矩阵向量乘运算的数据结构和计算过程,得到计算中不同数据的访存特征,并提出了一种面向数据访存特性的Cache划分方法。对12个稀疏矩阵向量乘的测试表明,本文的Cache划分方法能有效地提高可重用向量的Cache命中率,同时减少计算对Cache空间的需求。

Memory Access Behavior Characteristics-Oriented Cache Partition for SPMV

The sparse matrix vector multiplication ( SPMV ) is the kernel of many scientific applications.Due to its indirect and random memory accesses , the SPMV is the bottleneck of scientific applications.In this paper , based on analyzing data structure and computation process , we present the memory access characteristics of the SPMV and propose a memory access behavior characteristics-orient cache partition method.According to the performance evaluation with 12sparse matrix vector multiplications , the proposed cache partition method increase the hit ratio of reusable data and reduce the cache capacity request efficiently during the SPMV.