Improving Energy Consumption in Iterative Problems Using Machine Learning

Abstract

To reach the new milestone in High Performance Computing, energy and power constraints have to be considered. Optimal workload distributions are necessary in heterogeneous architectures to avoid inefficient usage of computational resources. Static load balancing techniques are not able to provide optimal workload distributions for problems of irregular nature. On the other hand, dynamic load balancing algorithms are coerced by energy metrics that are usually slow and difficult to obtain. We present a methodology based on Machine Learning to perform dynamic load balancing in iterative problems. Machine Learning models are trained using data acquired during previous executions. We compare this new approach to two dynamic workload balancing techniques already proven in the literature. Inference times for the Machine Learning models are fast enough to be applied in this environment. These new predictive models further improve the workload distribution, reducing the energy consumption of iterative problems.