Multivariate time series ELM for cloud data centre workload prediction

Abstract

In existing Cloud Data Centres (CDCs), workload prediction plays an important role in energy conservation, as it allows for dynamic migration and consolidation of Virtual Machines (VMs) which are provisioned on these centres. In this paper, we propose a new multivariate time series Extreme Leaning Machine (ELM) algorithm, and use it in an efficient CDC workload prediction framework based on energy consumption. This prediction framework not only uses VM historical usage values, but also takes into account VM and user behaviour and current states of the data centre. We introduce a number of techniques to handle the problem of predicting window sizes to optimize Physical Machine (PM) utilization. The proposed ELM algorithm and prediction framework are implemented using Google Trace data, which represents a 29-day trace collected from a cluster that contains more than 12,500 PMs. The results indicate that our model performs better on a variety of time series patterns than other models in the literature.