CPUs Energy Consumption Reduction for Asynchronous Parallel Methods Running over Grids

This paper presents a new online frequency selecting algorithm for asynchronous parallel methods running over grids. It selects a vector of frequencies that gives the best tradeoff between energy consumption and performance. It also has a very small overhead and works without training and profiling. New energy and performance models are used in this algorithm to predict the execution time and the energy consumption of synchronous, asynchronous and hybrid iterative applications running over grids. The algorithm was evaluated on the Grid'5000 testbed while running a multi-splitting application that solves a 3D problem. The experiments show that applying synchronously the proposed frequency scaling algorithm to the asynchronous version of the application reduces its energy consumption up to 26.93% and speeds it up by 21.48%. Finally, the algorithm was compared to the Energy and Delay Product (EDP) method and it outperformed the latter in the energy reduction and performance trade-off for parallel asynchronous iterative methods.