基于机群系统的不同并行I/O策略性能的研究

Performance of Parallel I/O Strategies Based on Cluster

基于并行程序的2种I/O访问策略,设计实现了高维矩阵-矩阵相乘和高维矩阵-向量相乘的并行算法,解决了内存或Cache中映射的数据对I/O测试结果的影响。实验结果表明:并行读取策略的I/O性能优于主节点读取策略。

We study two different I/O access strategies for parallel computation implemented on cluster systems,and develop parallel algorithms for high-dimensional matrix-matrix and matrix-vector multiplications.We present an effective solution to eliminating mapped experimental data in memory or Cache,which have impact on the I/O test results.Experiments show that the parallel read（PR） strategy of I/O is superior to the master read（MR） strategy.