Prediction based Auto Scaling for Cloud Applications

ABSTRACT Objectives: Cloud computing provides highly scalable environment to applications. Due to dynamicity of application workload, static allocation of resources is not suitable. Scalability provides a way to increase or decrease resources of particular application at runtime. Static allocation of resources may lead to several problems like over provisioning and under provisioning. If resources are allocated statically by considering peak workload (over provisioning), majority of time, resources will be underutilized and users will have to pay more. If allocation is done based on average workload, it will not be possible to achieve performance objective while peak load (Under provisioning). Horizontal scalability is provided by means of adding or removing VM instances where as, Vertical scalability can be achieved by increasing or decreasing resources allotted to particular VM. Automatic scaling of resources based on workload change reduces provisioning costs and helps clients to achieve performance objectives. Methods: Automatic scaling can be done either by reactive way or by proactive way. In reactive auto scaling, resources are scaled as a reaction of some event. In our work, proactive auto scaling method is suggested based on application workload prediction. Time series analysis based ARIMA model is used for workload prediction. Findings: Proactive auto scaling increases or decreases application resources in advance based on prediction. This helps to improve application performance as well as to minimize SLA violations. It also reduces unnecessary cost incurred to cloud user. Application/Improvement: ARIMA based time series prediction is used to predict future work load. Scaling decision is taken based on predicted future workload as well as current application average response time. Keywords: Server consolidation, SLA, Response time, ARIMA, Scalability, Cloud Computing, DCSim