大规模稀疏矩阵在并行应用中的通信优化研究

Research on optimization of large-scale sparse matrix communication in parallel application

为了提高并行应用系统的效率,研究了针对大型稀疏矩阵的压缩通信问题。通过对矩阵压缩通信过程中矩阵稀疏度、网络带宽、处理器计算能力之间的关系进行定量分析,推导出稀疏度下界计算公式。通过对不同稀疏度情况下算法所取得的效率进行分析,总结出压缩通信中稀疏度与通信效率之间的函数关系。结合油藏数值模拟的应用实例,设计实现了稀疏矩阵的压缩通信算法。结果表明本算法在稀疏矩阵通信方面效率有明显的提高。

This paper presented a method of compressed communication for sparse matrix in order to improve the performance of communication, obtained a formula for calculating the lower bound of sparsity by quantitative analysis of the relationship be- tween sparsity, network bandwidth and frequency of CPU. In addition,derived a function between sparsity and communication performance from analysis of the algorithm efficiency on different sparsities. Based on the application of parallel petroleum reservoir simulation ,implemented an algorithm for compressed communication. The experiments show that the performance of sparse matrix communication can be improved significantly.