Queuing Theory and Dynamic Programming to Model Resources Allocation in a Cloud Computing Environment

Abstract

Cloud computing has become a dominant paradigm in the IT environment in recent years. Its principle is to provide services to customers on demand, and allows them to pay only for what they need. The increasing need for this type of service brings the providers to increase the size of their infrastructure, as energy consumption and associated costs become very significant. A cloud service provider must provide all the different types of requests. Infrastructure managers need then to host all the types of services together. That's why during this thesis, we focus on the modeling and analysis of management resources in data center environments in cloud computing. Analytical models are developed to improve the quality of services deployed in the cloud and minimize the energy consumption. Two great analytical tools have been used to achieve this purpose, queuing theory and dynamic programming. To achieve this, we first modeled and studied the problem of initial resource allocation, proposing two algorithms approached via heuristics, and then exploiting the concept of virtual machine migration in a cloud data center. The notion of reserve servers is implemented in our models in order to manage energy consumption. Servers are divided into two types: In run mode and in sleep mode as a reserve when there is no work to be done. Usually, the service receives customers in one way or another, we have focused on M/M/K queue system with a setup cost that describes a delay when a new reserve server is turning on. In another approach, a controller is placed in front of the arriving clients to manage their placement in the system and he is the one who decides whether to turn on a reserved server or not depending on the client's needs. Typically, services have to deal with different load levels. This is why we have developed the resource allocation models including the dynamical nature of requests and resource usage, as well as the notion of the cost of switching servers ON and OFF and the cost of migrating services from one server to another. The results of the models used during the works have been validated by both analytical and numerical solutions and some of them are compared qualitatively with other recent works.