Abstract
Communication takes a significant amount of elapsed execution time in parallel CFD. Recent studies show that the high communication rates in most parallel CFD algorithms are necessary only for maintaining the stability of numerical integration, but not for solving the CFD problem. Once the stability is achieved using filtering techniques, a reduction of communication rate can significantly reduce the elapsed execution time. This study is to determine the optimal communication rates for all interfaces among parallel processes during the execution of parallel CFD. The objective is to minimize the overall elapsed program-execution time. How communication rates affect the elapsed execution time for parallel CFD applications is investigated first. Experiments demonstrate that the selection of communication rates based on human intuition is very difficult and unreliable. As the search of optimal communication rate in parallel CFD is an NP-hard problem, genetic algorithms are adopted for finding the suboptimal solutions. An automatic online optimization process for the selection of the optimal communication rate is developed.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call