Latency hiding by redundant processing
Jeremy F. Villalobos, Barry Wilkinson · 2008
The increase in interconnected computational resources brought about by the Grid creates the possibility to port multiple parallel programming techniques to it. Porting parallel applications to the Grid could reduce the total computation time, or it could be used to create solutions with higher degrees of resolution. However, the Grid brings with it network state conditions that all too often work to the detriment of expediency in parallel applications. This paper proposes an algorithm designed to significantly reduce the amount of Wide Area Network (WAN) latency experienced when running an interactive synchronous parallel program on the Grid. The algorithm is called Latency Hiding by Redundant Processing (LHRP) and in tests done on two Grid nodes with emulated latency, it complemented Latency Hiding (LH) by performing better than LH on jobs with low internal computation time and performing worst than LH when the computation time was enough to hide the latency.