Towards SLA-Driven Autoscaling of Cloud Distributed Services for Mobile Communications

Abstract

In recent years cloud computing has established itself as the computing paradigm that supports most distributed systems, which are essential in mobile communications, such as publish-subscribe (pub/sub) systems or complex event processing (CEP). The cornerstone of cloud computing is elasticity, and today’s autoscaling systems leverage that property by making scaling decisions based on estimates of future workload to satisfy service level agreements (SLAs). However, these autoscaling systems are not generic enough, as the workload definition is application-based. On the other hand, the workload prediction needs to be mapped in terms of SLA parameters, which introduces a double prediction problem. This work presents an empirical study on the relationship between different types of workloads in the literature and their relationship in terms of SLA parameters in the context of mobile communications. In addition, more than 30 prediction models have been trained using different techniques (time series analysis, regression, random forests) to test which ones offer better prediction results of the SLA parameters based on the type of workload and the prediction horizon. Finally, a series of conclusions on the predictive models to be used as a first step towards an autonomous decision system are presented.