Latency tolerant algorithms for WAN based workstation clusters

Abstract

One of the biggest differences between traditional supercomputers and workstation clusters is the latency involved in sending a message between processors. Wide Area Network (WAN) based workstation clusters can experience significant latency between machines at different geographical positions. Improvements in network technology can achieve marginal improvements, but the speed of light delays cannot be decreased. This research develops stencil algorithms that are more tolerant of latency. These algorithms can be used to solve finite element problems as well as other problems where neighbor communications are used. Latency tolerant algorithms are essential if a large number of machines on the Internet are to be used in performing a parallel computation.