Resumen
Keeping energy costs in budget and operating within available capacities of power distribution and cooling systems is becoming an important requirement for High Performance Computing (HPC) data centers. It is even more important when considering the estimated power requirements for Exascale computing. Power and energy capping are two of emerging techniques aimed towards controlling and efficient budgeting of power and energy consumption within the data center. Implementation of both techniques requires a knowledge of, potentially unknown, power and energy consumption data of the given parallel HPC applications for different numbers of compute servers (nodes).This paper introduces an Adaptive Energy and Power Consumption Prediction (AEPCP) model capable of predicting the power and energy consumption of parallel HPC applications for different number of compute nodes. The suggested model is application specific and describes the behavior of power and energy with respect to the number of utilized compute nodes, taking as an input the available history power/energy data of an application. It provides a generic solution that can be used for each application but it produces an application specific result. The AEPCP model allows for ahead of time power and energy consumption prediction and adapts with each additional execution of the application improving the associated prediction accuracy. The model does not require any application code instrumentation and does not introduce any application performance degradation. Thus it is a high level application energy and power consumption prediction model. The validity and the applicability of the suggested AEPCP model is shown in this paper through the empirical results achieved using two application-benchmarks on the SuperMUC HPC system (the 10th fastest supercomputer in the world, according to Top500 November 2013 rankings) deployed at Leibniz Supercomputing Centre.