Performance analysis for noncontiguous I/O using multiple I/O, data sieving and list I/O in peer-to-peer parallel computing

As far as P2P high performance science computing is concerned, one of the many challenges is to provide effective ways to access noncontiguous file data. Three noncontiguous file access strategies used in traditional parallel file system, including multiple I/O, data sieving and list I/O, are analyzed. It is pointed out that routing is needed in P2P environment to locate the data server, while it needs no server locating in a parallel file system of a cluster. It is found that the file segments are more disperse in a P2P network than in a cluster. After figuring out the processes of these three different noncontiguous I/O strategies, the time models are set up. Simulation is carried out on that time models, with various client number, trunk size and access size. As campus network is concerned, it shows that data-sieving is inferior to others. Multiple I/O and list I/O have similar performance for data access with large access size.