Abstract
There is a trend in building a system by interconnecting non-dedicated computing nodes through public wired or wireless networks. Unstable communication latency and unpredictable CPU utilization can be experienced by the users on such a network computing platform. The adverse effects can be avoided by hiding communication latency. However, we have shown that communication latency can not be effectively hidden without carefully choosing a proper message scheduling strategy. In this paper, two parameters are proposed to model the network computing platform. Based on the model, a function can be defined to capture the message scheduling strategy of an algorithm. Our message scheduling paradigm can be employed to develop high performance programs. Finally, matrix multiplication and LU decomposition are used as examples to illustrate the usefulness of our paradigm.