Prédiction dynamique de la consommation de ressources dans des environnements virtualisés

Abstract

Cloud computing environments are in a continual progress and offer a growing number of capacities to develop and to host new applications and services. In these environments, managing efficiently resources is considered as one of many and major challenges to face in order to respect negotiated service contracts (Service Level Agreement: SLA) required to implement these applications/services. In fact, the variability and unpredictability of traffic loads and Quality requirements Of Services (QoS) have a direct impact on the resource consumption. Thus, a performant resources management not only must guarantee the services quality requirements defined in the SLA contract to avoid the cost induced by their violation, but also an effective consumption of shared resources. Moreover, unusual and sudden changes in the traffic load, which are created by emergency cases (e.g.: software update to consolidate the infrastructures security) or breakdowns, produce a serious services performance degradation and costly SLA violations. In this context, the solutions that are based on the Cloud infrastructures resizing, resources consumption prediction methods using the thresholds or statistic models (e.g.: linear regression manipulating data retrieved from an offline historic) are inadequate. Considering this finding, we propose in our thesis a new solution that apply machine learning to manage efficiently resources in Cloud environments. Our solution is characterized by : a dynamic and reliable resources consumption prediction applicable to many systems ; automatic and adaptive tuning of the prediction combined with automatic adjustment of the sliding-window size of training data/predicted data (optimal size); an abnormal variations detection automatically and a mapping between metrics at service and resource levels plus a detection of their periodicity; automatic system state assessment (e.g. : a state update after a normal variation caused by resources adaptation or an abnormal change generated by a hardware/software failure in an infrastructure) and notification generation. Algorithms and technics that we propose have to optimize the resources sharing, to minimize consumption costs and to satisfy SLA requirements. The analysis of our experimental tests results show clearly the relevance and efficiency of our solutions.

Date	16 Dec 2020
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Nadjia Kara (Supervisor)