Machine Learning based Load-Balancing for the CESM Climate Modeling Package

TitleMachine Learning based Load-Balancing for the CESM Climate Modeling Package
Publication TypeReport
Year of Publication2013
AuthorsBalaprakash, P, Alexeev, Y, Mickelson, SA, Leyffer, S, Jacob, RL, Craig, AP
CityDenver, CO
Other NumbersANL/MCS-P4070-0413
Abstract

Our goal is to find optimal load-balancing parameter configurations for the Community Earth System Model (CESM) on Intrepid IBM BlueGene/P. We developed an effective machine-learning based load-balancing algorithm. The novelty of the algorithm is in using a predictive model to bias the search towards high performing load-balancing parameter configurations, that are iteratively evaluated on the target architecture and reused to improve the accuracy of the predictive model on promising regions of the search space. Experimental results show that the proposed algorithm can significantly reduce time and core-hours usage required for finding high quality parameter configurations without a significant loss in performance. When compared to the current practice of complete enumeration over feasible configurations, the machine-learning based load-balancing algorithm requires six times fewer evaluations to find configurations, whose quality is only 2% worse than the optimal configurations.

 

PDFhttp://www.mcs.anl.gov/papers/P4070-0413.pdf