Melhoria da eficiência energética de comitês de classificadores de fluxo de dados para computação de borda

Abstract

Edge computing (EC) has emerged as an architecture that can help reduce the energy demand and greenhouse gas emissions of digital technologies. Edge computing offers low latency, mobility, and location awareness for delay-sensitive devices, connecting cloud computing services to end-users. Machine learning (ML) methods have increasingly been used on edge devices for data classification and information processing. Classifier ensembles have demonstrated good predictive performance in data stream classification problems. The strategy called mini-batching was previously proposed in the literature to improve cache data reuse when executing classifier ensembles on multi-core architectures for online data stream classification. The strategy involves temporarily grouping data from a data stream and processing them together. As a result, mini-batching can speed up applications and reduce energy consumption. However, the originally proposed mini-batching offers opportunities for further improvements. In this work, we investigate the fusion of the training and classification stages of the data, bringing more gains in cache reuse and predictive performance improvements. We also evaluate the mini-batching strategy compared to two strategies supported by the hardware of common multi-core processors used in edge devices: clock frequency reduction and processor core shutdown. We evaluate the strategies by comparing their performance and energy efficiency for data stream classification using six state-of-the-art classifier ensemble algorithms and four benchmark datasets. The results show that mini-batching strategies can significantly reduce energy consumption in 95% of the experiments, improving energy efficiency by an average of 96% and by 169% in the best case over hardware strategies. Similarly, the newly proposed mini-batching strategy improved energy efficiency by an average of 136% and 456% in the best case. Finally, we proposed an adaptive and multi-objective optimization strategy to dynamically choose the mini-batching size based on CPU occupancy and data arrival rates. The batch size choice uses the Pareto principle to optimize both response time and energy consumption. Results show an improvement in energy consumption in 17 of the 24 cases evaluated. However, for the latency metric, there was no significant reduction compared to batch sizes of 50 (pointed out in the literature as a good choice). In summary, the dynamic strategy offers reduced energy consumption without losses in execution time.