Optimization of the Tradeoffs

As the measurement and analytical results indicate, if the number of client nodes is smaller than the number of server nodes, server-driven protocols tend to have a higher write performance than the client-driven protocols since more nodes are involved in sharing the duplication work. Between the server driven protocols, the synchronous one is preferred because it has a higher reliability with only slightly lower bandwidth. On the contrary, if the total number of the client nodes is greater than that of the server nodes, the client-driven protocols are better than their server-driven counterparts. Between the client-driven protocols, the asynchronous client duplication is the most favorable since it has the highest write performance and the second best reliability. These observations lead us to propose a hybrid protocol to optimize the tradeoff between the reliability and bandwidth performances.

A scientific application is usually required to specify the total number of parallel jobs or clients it needs before running in a cluster. In the hybrid duplication protocol, each client compares the total server number in one storage group with the total number of parallel clients of the current application. If the server number exceeds the client number, the synchronous server

duplication is used to mirror the data. Otherwise, the asynchronous client duplication is preferred. In this way, this hybrid protocol always tries to achieve a considerably high write performance or reliability with little degradation of the other.