Utility-based Allocation of Resources to Virtual Machines in Cloud Computing

In recent years, cloud computing has gained a wide spread use as a new computing model that offers elastic resources on demand, in a pay-as-you-go fashion. One important goal of a cloud provider is dynamic allocation of Virtual Machines (VMs) according to workload changes in order to keep application performance to Service Level Agreement (SLA) levels, while reducing resource costs. The problem is to find an adequate trade-off between the two conflicting objectives of application performance and resource costs. In this dissertation, resource allocation solutions for this trade-off are proposed by expressing application performance and resource costs in a utility function. The proposed solutions allocate VM resources at the global data center level and at the local physical machine level by optimizing the utility function. The utility function, given as the difference between performance and costs, represents the profit of the cloud provider and offers the possibility to capture in a flexible and natural way the performance-cost trade-off. For global level resource allocation, a two-tier resource management solution is developed. In the first tier, local node controllers are located that dynamically allocate resource shares to VMs, so to maximize a local node utility function. In the second tier, there is a global controller that makes VM live migration decisions in order to maximize a global utility function. Experimental results show that optimizing the global utility function by changing the number of physical nodes according to workload maintains the performance at acceptable levels while reducing costs. To allocate multiple resources at the local physical machine level, a solution based on feed-back control theory and utility function optimization is proposed. This dynamically allocates shares to multiple resources of VMs such as CPU, memory, disk and network I/O bandwidth. In addressing the complex non-linearities that exist in shared virtualized infrastructures between VM performance and resource allocations, a solution is proposed that allocates VM resources to optimize a utility function based on application performance and power modelling. An Artificial Neural Network (ANN) is used to build an on- line model of the relationships between VM resource allocations and application performance, and another one between VM resource allocations and physical machine power. To cope with large utility optimization times in the case of an increased number of VMs, a distributed resource manager is proposed. It consists of several ANNs, each responsible for modelling and resource allocation of one VM, while exchanging information with other ANNs for coordinating resource allocations. Experiments, in simulated and realistic environments, show that the distributed ANN resource manager achieves better performance-power trade-offs than a centralized version and a distributed non-coordinated resource manager. To deal with the difficulty of building an accurate online application model and long model adaptation time, a solution that offers model-free resource management based on fuzzy control is proposed. It optimizes a utility function based on a hill-climbing search heuristic implemented as fuzzy rules. To cope with long utility optimization time in the case of an increased number of VMs, a multi-agent fuzzy controller is developed where each agent, in parallel with others, optimizes its own local utility function. The fuzzy control approach eliminates the need to build a model beforehand and provides a robust solution even for noisy measurements. Experimental results show that the multi-agent fuzzy controller performs better in terms of utility value than a centralized fuzzy control version and a state-of-the-art adaptive optimal control approach, especially for an increased number of VMs. Finally, to address some of the problems of reactive VM resource allocation approaches, a proactive resource allocation solution is proposed. This approach decides on VM resource allocations based on resource demand prediction, using a machine learning technique called Support Vector Machine (SVM). To deal with interdependencies between VMs of the same multi-tier application, cross- correlation demand prediction of multiple resource usage time series of all VMs of the multi-tier application is applied. As experiments show, this results in improved prediction accuracy and application performance