Network Performance-aware Communication Model for Clustered Wide Area Systems
Yanhui Wu · Jisuanji gongcheng · 2004
Multiple clusters in computational grid are connected via wide area networks. Hierarchical network structure is a major problem in parallel computing applications for such platforms: latency and bandwidth of wide area networks are often orders of magnitude worse than those of local networks. To optimize graph shape and segment size at runtime, the paper introduces a performance model called parameterized LogP, a hierarchical extension of the LogP model that covers messages of arbitrary length, and discusses all parameters and how to get them through experiments. With selecting suitable communication graph shapes, and splitting messages into multiple segments that are sent in parallel over different WAN links based on new communication model to reach communication optimization.