Análise do desempenho e consumo de energia de coprocessadores xeon phi utilizando diferentes modelos de programação

Abstract

With computational growth occurring in the last few years, issues and problems related to energy costs arise mainly in large computing environments such as data centers or high-performance processing centers because they require high energy demand to maintain their operation. In the area of High Performance Computing (HPC), following this need arise new computational paradigms to support the exponential increase of processor cores, changing from models and multicore architectures where they centralized the processing only in the CPUs for new heterogeneous architectures and manycore that distribute processing using GPUs and coprocessors. The performance relationship with efficient energy consumption in manycore architectures is still a recent issue and there are many questions that need to be explored, because along with the increase in the number of colors have appeared several programming models. Greater study on power consumption is required in models that use and manage part workloads on CPUs and part on GPUs or coprocessors. In this context, the general objective of this work is to perform a study and methodology on the relation between performance and energy consumption in Intel Xeon Phi coprocessors, analyzing and identifying the main factors that directly affect the efficient consumption of energy in environments manycore . The evaluation will be based on the behavior and variation of the energy consumption during the execution of the LINPACK, HPL 2.1 and HPCG benchmarks on the programming languages host, offload, native and symmetric proposed by Intel. The methodology used consists of scenarios planning, execution, monitoring, collection, measurement and analysis of energy consumption data and performance on the results generated, mainly the definitions of the evaluated criteria, the metrics used and the scenario models used. The main contributions of this work are: Evaluate which programming models provide better results with efficient consumption without compromising performance; Analyze and identify if the behavior of shared memory resources and features, number of nodes, cores, processes and threads directly impact the performance and power consumption relationship. Results show that there is great variation in the performance and energy consumption between the models and the importance of the choice and proper configuration of the factors used during the parallel application processing. Considering shared memory scenarios on the Linpack benchmark, the host and offload models present a linear increase in performance for workloads (size) up to 10000 and a slight increase after this value. But the native model in spite of lower performance in these scenarios, in contrast consumes less energy and mainly more favorable for applications with a high degree of parallelism.