Détection et analyse de données aberrantes pour l'adaptation de ressources dans des environnements virtualisés

Abstract

The integration of the cloud computing with telecommunications technologies has brought several benefits to network operators and communication service providers. Indeed, the use of several virtual nodes of a service platform deployed on the same physical machine can significantly reduce costs and energy consumption. With the elasticity and scalability offered by this new technology, controlling workloads in virtualized systems has become a necessity to ensure a good quality of service. In this research, we discuss the problem of workload management in virtualized environments. Our goal is to propose and validate an adaptation mechanism that ensures a dynamic and efficient resource management in virtualized environments. The proposed approach consists in the first place of testing and analyzing the performance of a virtualized IMS platform that we installed based on OpenIMS Core. Then, we analyzed the data collected from IMS to detect potential outliers. In addition, we performed the mapping between service-level and resource-level metrics through the analysis of their variation using the modified z-score method. This analysis was also based on data extracted from other virtualized environments such as Google Cluster and the ÉTS in order to have a generic solution adapted to several virtualized systems. In addition, we performed a comparison between the two methods of outliers detection namely modified z-score and mahalanobis. Our analysis showed that mahalanobis gives us better results compared to modified z-score. The results obtained made it possible to identify important variations that may require the adaptation of resources in the system. Finally, we designed and developed a resource adaptation algorithm based on the Mahalanobis method.

Date	11 Dec 2018
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Nadjia Kara (Supervisor)