Scaling Policies Derivation for Predictive Autoscaling of Cloud Applications
Yesika Marlen Ramirez Cardenas · mediaTUM – the media and publications repository of the Technical University Munich (Technical University Munich) · 2018
Cloud Computing enables the provision of services based on a pay-as-you-go model.Allocation and release of resources happen dynamically according to customers' needs following a Service Level Agreement (SLA).However, as there are a wide diversity of cloud applications and different business goals, the selection of a provisioning plan that adapts to the business dynamic becomes challenging.Cloud applications may have an under-provisioning or over-provisioning problem since the Cloud Service Provider face the trade-off between minimize resources costs and meet the stringent Quality of Service (QoS) requirements.Known Auto Scaling solutions adjust the number of available computing resources according to the current service demand but they are either restricted to a single type of resource or they do not consider the type of cloud application to be scaled and do not anticipate workloads that arise at runtime which may cause SLA violations.This thesis aims to improve the Auto Scaling process by deriving a set of scaling policies through the analysis of load patterns, application's performance profiles, resource types, and pricing model.The outcome of this research includes the implementation of the Scaling Policy Derivation Tool (SPDT) which based on different scaling strategies selects a cost-efficient scaling policy able to meet the current and predicted service demand.The proposed solution helps Cloud Services Providers to efficiently manage the deployment and scaling of applications by answering questions like Which is the right number of resources needed?, what type or combination of type resources fits better?, what parameters and conditions should be considered before scaling?, When to start the scaling of resources?, how the type of application influence the scaling decisions?.