SLA based Service Differentiation in Cloud Data Center

In this thesis, we investigate a problem in the realm of hosting Internet of Things (IoT) broker services in a cloud computing datacenter. IoT sensors generate measurements that are transported to the datacenter for processing. Several tenants, each with his/her own measurement generation patterns are housed in such a center. The tenants and datacenter provider enter a Service Level Agreement (SLA) that formally states the service expectations (tenant) and obligations (provider). In the particular SLA we have considered, the only expectation from the tenant is to "have a specific amount of measurements processed within a given time period"; the provider's obligation is to provide the CPU resources to meet this expectation regardless of the details of the measurement generation pattern. The main merit of this SLA is its (technical and legal) simplicity.;The main technical problem we address is how can this SLA be guaranteed? This is a question that the datacenter designer must answer. The designer has several tools at his/her disposal in answering this question - we call such tools "knobs" that can be introduced in algorithmic solutions. The main knob we focus on, in this work, is CPU time provisioning.;We tackle this technical problem by breaking it down to three smaller problems. First, we study the SLA in its present form.We formulate two problems; in the first, we assume that a single server can provide the needed time to meet the tenant expectation. The control knob is fairly simple to design in this environment. In the second, we consider the situation where several servers must be provisioned, in order to meet the expectation. In the third formulation of the problem, we modify the SLA -we remove some of its simplicity- in order to simplify the designed controls.;The main technical challenge in all problems when deciding how to provision CPU time is the lack of knowledge about traffic patterns. An additional challenge is present in the case of multiple servers - the presence of two conflicting objectives that come from the tenant and the cloud provider. The tenant objective can be expressed in terms of, say, acceptable average delays, while the provider's objective can be expressed in terms of, say, acceptable operating cost. Even though these objectives are not explicitly stated in the SLAs, it is to the best interest of all parties involved if the designer was to come up with resource provisioning controls that will keep both parties happy.;The main contribution of this thesis is twofold: (a) the introduction of the new SLA, and, (b) the design and evaluation of the CPU provisioning algorithms to meet the SLAs.