Stability Analysis of a Statistical Model for Cloud Resource Management

In this paper, we presented a comprehensive stability analysis of statistical models derived from the network usage data to design an efficient and optimal resource management in a Cloud data centre. In recent years, it has been noticed that network has a significant impact on the HPC and business critical applications when they are run in a cloud environment. The existing VM placement algorithms lack capabilities to deploy such applications in an effective way and cause performance degradation. As a result, there is an urge for a network-aware VM placement algorithm which will consider the application behaviour and system capability. Our approach uses static models based on simple probability distribution concept and partition (number theory) to characterise and predict the resource usage behaviour of the VMs. However, the stability of those models is a key requirement to ensure a persistent placement of the VMs which can prevent their frequent migration and keep the infrastructure rigid. The paper investigates the stability of the models with respect to time. Sticky HDP-HMM method was proven highly capable to model the monitoring data with a certain accuracy. The refined data was further used to estimate the resource consumption of each VM and physical host running in the infrastructure. A stability parameter has been defined to determine the level of steadiness of the models that gives us a clear indication on whether the models can be used further to derive an optimal placement decision for new VMs. The paper ends with a discussion on instance based stability analysis and future work.