LogGP(h): Incorporating Communication Hierarchy into the LogGP Model
In a large-scale cluster system, thousands of computing nodes increase the complexity of network topology. Nonetheless, there are few existing point-to-point communication models considering the impact of hierarchical communication latencies and bandwidths caused by the net work complexity in such a large system. We propose a new parallel com putational model called LogGP(h) with a new parameter h incorporated into LogGP model to abstract communication hierarchy. By predicting and analyzing the point-to-point and collective communication under an 8-blade IBM JS21 cluster with this new model, we find LogGP(h) is more practical than LogGP model.
Yuxin Tang Yunquan Zhang Xiangzheng Sun
Laboratory of Parallel Computing,Institute of Software,Chinese Academy of Sciences State Key Laborat Laboratory of Parallel Computing,Institute of Software,Chinese Academy of Sciences State Key Laborat
国际会议
南宁
英文
96-108
2009-12-04（万方平台首次上网日期，不代表论文的发表时间）