多核处理器机群点对点通信同时具有memory纵向层次化特征和横向层次化的新特征.纵向层次化特征揭示了对不同大小和步长的消息进行点对点通信时消息通信中间件对其性能的影响;横向层次化的新特征由intra-CMPi、nter-CMP和inter-node消息通信性能的显著差异引起,目前缺少有效的分析模型.文中提出一种新的memory层次化并行计算模型,对多核处理器机群memory横向、纵向层次化特征进行了统一的抽象.在对多核处理器机群点对点通信和集合通信的开销进行模型分析和实际测试中,新模型的精确性优于现有的未引入memory横向层次化特征的模型.
Point-to-point message communication on multi-core clusters has the feature of memory hierarchy not only in the vertical level,but in the horizontal level.Memory hierarchy in the vertical level may be used to analyze middleware′s effects on point-to-point communication with different message sizes and message strides;Memory hierarchy in the horizontal level is brought by distinct message communication performance among intra-CMP,inter-CMP and inter-node,which is lack of effective analysis model.In this paper,new parallel computation model are proposed to abstract memory hierarchy in vertical and horizontal levels,and the overhead of point-to-point and collective communication under new model are evaluated.The results of performance evaluation show that new model is more suitable for multi-core cluster and facilitated the optimization of MPI middleware and parallel applications on multi-core clusters.