Research on identification of server energy consumption characteristics via dirichlet max-margin factor analysis similarity preservation model

Chen, Buhua and Liu, Hanjiang and Shen, Chengbin and Shen, Buyang and Li, Kunlun (2023) Research on identification of server energy consumption characteristics via dirichlet max-margin factor analysis similarity preservation model. Frontiers in Energy Research, 10. ISSN 2296-598X

[thumbnail of pubmed-zip/versions/2/package-entries/fenrg-10-1064464-r1/fenrg-10-1064464.pdf] Text
pubmed-zip/versions/2/package-entries/fenrg-10-1064464-r1/fenrg-10-1064464.pdf - Published Version

Download (4MB)

Abstract

Growing server energy consumption is a significant environmental issue, and mitigating it is a key technological challenge. Application-level energy minimization strategies depend on accurate modeling of energy consumption during an application’s execution. This paper presents a theoretical and experimental study of the dpMMSPFA model in the field of server energy consumption identification. The dpMMSPFA for classification of hidden spaces uses latent variable support vector machines (LVSVM) to learn discriminative subspaces with maximal marginal constraints. The factor analysis (FA) model, the similarity preservation (SP) item, the Dirichlet process mixture (DPM) model, and the maximal marginal classifier are jointly learned beneath a unified Bayesian architecture to advance classification of predictive power. The parameters of the proposed model can be inferred by the simple and efficient Gibbs sampling in terms of the conditional conjugate property. Empirical results on various datasets demonstrate that 1) max-margin joint learning can significantly improve the prediction performance of the model implemented by feature extraction and classification separately and meanwhile retain the generative ability; 2) dpMMSPFA is superior to MMFA when employing SP item and Dirichlet process mixture as prior knowledge; 3) the classification of dpMMSPFA model can often achieve better results on benchmark and measured energy server consumption datasets; 4) and the recognition rate can reach as high as 95.79% at 10 components, far better than other models on measured energy server consumption datasets.

Item Type: Article
Subjects: Article Archives > Energy
Depositing User: Unnamed user with email support@articlearchives.org
Date Deposited: 01 May 2023 05:57
Last Modified: 19 Jun 2024 11:51
URI: http://archive.paparesearch.co.in/id/eprint/1188

Actions (login required)

View Item
View Item