Decentralized Multi-Resource Allocation in Clouds

Cloud computing is omnipresent nowadays, as it allows for a fine-grained partitioning of data center resources and for flexibly providing access to these resources. This partitioning and provisioning is achieved by hosting several Virtual Machines (VM) on the same physical machine. In contrast to commercial clouds, the performance of VMs in a private cloud is not captured by Service Level Agreements, and thus, all VMs are treated as processes of equal importance. As users operate...
More Description

Cloud computing is omnipresent nowadays, as it allows for a fine-grained partitioning of data center resources and for flexibly providing access to these resources. This partitioning and provisioning is achieved by hosting several Virtual Machines (VM) on the same physical machine. In contrast to commercial clouds, the performance of VMs in a private cloud is not captured by Service Level Agreements, and thus, all VMs are treated as processes of equal importance. As users operate different numbers of VMs and these VMs utilize different amounts of physical resources, this equal treatment of VMs leads to users receiving unequal amounts of physical resources. This thesis improves this situation by defining an efficient approach to enforce fairness in private clouds.This thesis shows that cloud resources are best controlled by changing priorities of VMs to access physical resources of their host and that no assumptions on utility functions can be made during this step. The premiss of this thesis that it is fair to constrain greedy users in favor of less greedy users requires a metric that quantifies the greediness of users based on their multi-resource self-servings from a shared resource pool. Thus, the Greediness Metric is developed based on a questionnaire among more than 600 participants on the intuitive understanding of greediness and fairness. The Greediness Metric is refined to define cloud fairness in a way that outperforms all existing cloud fairness definitions. To demonstrate the practical applicability of this cloud fairness definition, OpenStack is extended by an according service. The processing overhead of this service is evaluated and it is proven that it enforces fairness among users by coordinating the VM prioritization on hosts.