Performance and energy efficiency analysis of machine learning algorithms towards green al: a case study of decision tree algorithms

• Ano de defesa: 2021
• Autor(a) principal: Silva, Gabrieli Dutra
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Laboratório Nacional de Computação Científica; Coordenação de Pós-Graduação e Aperfeiçoamento (COPGA); Brasil; LNCC; Programa de Pós-Graduação em Modelagem Computacional
• Palavras-chave em Português: Computação de alto desempenho; Eficiência energética; Algoritmos (Computação); Aprendizado por computador; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO
• Link de acesso: https://tede.lncc.br/handle/tede/337

Artificial Intelligence (AI) is one of the most transformative forces today, achieving surprising results for solving problems in several application domains. These results are largely due to the use of the high computational capacity offered by HPC environments, which at the same time require a lot of energy for their functioning. Besides, energy consumption is responsible for greenhouse gas emissions, of which CO2 is the most significant. Machine Learning (ML) algorithms are generally evaluated and developed in terms of predictive performance. Little is known about its computational requirements and energy consumption, which has become an important task in achieving greener computing. The objective of this work is to evaluate the computational performance and energy efficiency of ML algorithms, specifically Decision Tree algorithms, to find solutions that can contribute to a GreenAI. Is evaluated the relations between: i) computational performance and energy efficiency when the algorithms are executed with different datasets, to analyze the influence of the number of examples and attributes on the performance and energy consumption; ii) energy efficiency and predictive performance, to assess the impact of the different parameter settings of the algorithms on energy consumption and, consequently, on the CO2 equivalent emission linked to their execution. Also, the parts of the code that consume the most energy are evaluated, what we call energy hotspots in this work.