Cloud Resources Forecasting based on Server Workload using ML Techniques

Abstract

Provisioning of resources, automatically, according to the service demand allows cloud service providers (CSPs) to implement elastic services. It is highly essential to guarantee QoS (Quality of Service) and to fulfill SLA (Service Level Agreement), in particular for the services having strict QoS needs like web servers getting a heavy load. Over-provisioning occurs when there is low demand, under-provisioning when there is a high demand and both of these situations result in poor quality of Experience (QoE) for the users. So, an approach is required to forecast future workload accurately prior to the resource allocation so that the right amount of resources can be provisioned. Such a system can help CSPs to achieve efficient resource allocation that maximizes their economic growth and also gains users’ satisfaction. This article presents an efficient workload forecasting mechanism built with the Support Vector Regression (SVR) technique to evaluate the number of resources required using web server workload time series data and queuing theory. The main goal of this technique is to forecast the future workload of the web server and predict the resources required to minimize the latency while reducing infrastructure costs and energy consumption. Normalization and outlier removal are performed for better forecasting. Various other techniques are also compared in estimating workload. Forecasting of load and prediction of resources is done with good accuracy by our proposed model better than many other models.