论文部分内容阅读
为了解决消息传递接口(MPI)的集合通信性能在实际应用中受非平衡进程到达模式影响较大的问题,提出了基于进程到达模式感知的广播优化方法.根据多核集群层级特点,在流水线广播算法基础上,优化算法以节点内与节点间两级感知进程到达,使先到达的进程尽快开始通信操作,减少等待时间,提高通信效率.理论分析模拟结果显示优化算法几乎不受非平衡进程到达模式影响.试验测试了在不同消息大小、不同进程到达模式和不同进程数三种情况下的算法性能,结果表明:本文所提出的算法有效提高了MPI广播在非平衡进程到达模式下的性能,通信时间减少129%~174%.
In order to solve the problem that aggregate communication performance of message passing interface (MPI) is greatly affected by unbalanced process arrival pattern in practice, a broadcast optimization method based on process arrival pattern awareness is proposed.According to the characteristics of multi-core cluster, Based on the optimization algorithm, the two-level perceptual process arrives in the node and the node, so that the first arriving process can start the communication operation as soon as possible to reduce the waiting time and improve the communication efficiency. The theoretical analysis shows that the optimization algorithm is hardly affected by the unbalanced process arrival pattern The experimental results show that the algorithm proposed in this paper can effectively improve the performance of MPI broadcasting in the non-equilibrium process arrival mode, and the communication The time is reduced by 129% ~ 174%.