Evaluation of synthetically generated traces towards a data-centre digital twin / Alejandro Fernández-Montes, Dámian Fernández-Cerero, Agnieszka Jakóbik, Belèn Bermejo, Carlos Juiz // W: ECMS 2023 ... więcej

Evaluation of synthetically generated traces towards a data-centre digital twin / Alejandro Fernández-Montes, Dámian Fernández-Cerero, Agnieszka Jakóbik, Belèn Bermejo, Carlos Juiz // W: ECMS 2023 : proceedings of the 37th ECMS International Conference on Modelling and Simulation, June 20th – June 23rd, 2023 Florence, Italy [Dokument elektroniczny] / ed. by Enrico Vicario [et al.] ; org. by ECMS - European Council for Modelling and Simulation. – Caserta : ECMS, 2023. – (Communications of the ECMS, ISSN 2522-2422 ; Vol. 37, Iss. 1). – S. 528-534 : tab., wykr. – Tryb dostępu: https://www.scs-europe.net/conf/ecms2023/ecms2023proceedings.pdf. – Bibliogr. 10 poz., Abstr. – Wydaw. wg cop. – Miejsce wyd. wg siedziby wydaw. – ISBN 978-3-937436-80-7 (Print). – ISBN 978-3-937436-79-1 (CD-ROM)

Several approaches exist to generate synthetic data centre traces for various purposes: from augmenting operational traces for data centre simulators and digital twins to forecasting incoming workload ... więcej

Several approaches exist to generate synthetic data centre traces for various purposes: from augmenting operational traces for data centre simulators and digital twins to forecasting incoming workload to improve data centre behaviour. The evaluation of the quality of synthetically generated multivariate time-series datasets, such as those related to data-centre traces, is not a trivial task, since complex patterns and correlation between variables may be present. This paper proposes a new multivariate time-series evaluation framework that computes a set of metrics and figures that can be used to measure the quality of synthetically generated data-centre traces. We then employ the proposed tool to compare two synthetic data centre traces with the original trace and assess their quality. These synthetic traces have been generated by means of Generative Adversarial Networks (GAN). In this work, we employ TimeGAN, a GAN model focused on the generation of multivariate time series traces. We finally show how the proposed framework provides us with a set of metrics consistent with the observable behaviour and numerical insights on the quality of the generated data centre traces, which are hard to acquire otherwise.