Energy Measurement and Modeling in High Performance Computing with Intel's RAPL

Significant advancements in the cloud computing paradigm have persuaded service providers to offer new and old services using the cloud computing platform for advantages like elasticity, scalability, availability and cost-effectiveness. In addition, the goal of achieving exaflops computation by 2020 by the High Performance Computing (HPC) community and the rapid growth in data generated and analyzed in the scientific computing paradigm have paved the way for an unprecedented growth in the number of server systems in data centers. As an example, CERN is now producing approximately 30 petabytes of data annually, which need to be stored and analyzed for particle physics. The proliferation of applications like social networking, video on demand and big data, is just adding more to the total number of server systems in data centers. Such big numbers of power hungry servers have increased the energy demand of data centers, and as a result energy efficiency in HPC, scientific computing and cloud computing is now a big concern. In this thesis, we investigate the energy consumption of server based computing systems and propose practical solutions for measuring, modeling and analyzing the energy efficiency of such systems. In this thesis, we have extensively used and analyzed Intel's Running Average Power Limit (RAPL) as an energy measurement tool. Firstly, we have used RAPL to profile the performance and energy consumption of an application. Secondly, we propose two strategies to model the power consumption of computing systems: modeling the power consumption of components inside the CPU such as instruction decoders, L2 and L3 caches, etc and modeling the full system power consumption using operating system counters and RAPL. For modeling the power consumption, we have used regression based models, statistical models as well as non-linear additive models. To validate our findings, we have used real production logs from data center as well as instances from Amazon Elastic Compute Cloud (EC2). The proposed power models predict the power consumption with promising accuracy. Thirdly, we have performed an extensive evaluation of RAPL as a power measurement tool and pinpointed RAPL's performance with respect to measurement overhead, accuracy, granularity, etc. This comprehensive analysis also reveals some open issues with RAPL that might weaken its usability in certain scenarios for which we also pinpoint solutions. Finally, to show the applicability of RAPL, we analyze the energy efficiency of two large scale graph processing platforms: Apache Giraph and Spark's GraphX.