Cloud computing is a multi billion dollar industry, based around outsourcing the provisioning and maintenance of computing resources. In particular, Infrastructure as a Service (IaaS) enables customers to purchase virtual machines in order to run arbitrary software. IaaS customers are given the option to purchase priority access, while providers choose whether customers are preempted based on priority level. The customer decision is based on their tolerance for preemption. However, this decision is a reaction to the provider choice of preemption policy and cost to purchase priority. In this work, a non-cooperative game is developed for an IaaS system offering resource reservations. An unobservable $M|G|1$ queue with priorities is used to model customer arrivals and service. Customers receive a potential priority from the provider, and choose between purchasing a reservation for that priority and accepting the lowest priority for no additional cost. Customers select the option which minimizes their total cost of waiting. This decision is based purely on statistics, as customers cannot communicate with each other. This work presents the impact of the provider preemption policy choice on the cost customers will pay for a reserved instance. A provider may implement a policy in which no customers are preempted (NP); a policy in which all customers are subject to preemption (PR); or a policy in which only the customers not making reservations are subject to preemption (HPR). It is shown that only the service load impacts the equilibrium possibilities in the NP and PR policies, but that the service variance is also a factor under the HPR policy. These factors impact the equilibrium possibilities associated to a given reservation cost. This work shows that the cost leading to a given equilibrium is greater under the HPR policy than under the NP or PR policies, implying greater incentive to purchase reservations. From this it is proven that a provider maximizes their potential revenue from customer reservations under an HPR policy. It is shown that this holds in general and under the constraint that the reservation cost must correspond to a unique equilibrium. / 2020-06-03T00:00:00Z

Network virtualization allows operators to host multiple client services over their base physical infrastructures. Today, this technique is being used to support a wide range of applications in cloud computing services, content distribution, large data backup, etc. Accordingly, many different algorithms have also been developed to achieve efficient mapping of client virtual network (VN) requests over physical topologies consisting of networking infrastructures and datacenter compute/storage resources. However as applications continue to expand, there is a growing need to implement scheduling capabilities for virtual network demands in order to improve network resource utilization and guarantee quality of service (QoS) support. Now the topic of advance reservation (AR) has been studied for the case of scheduling point-to-point connection demands. Namely, many different algorithms have been developed to support various reservation models and objectives. Nevertheless, few studies have looked at scheduling more complex "topology-level'' demands, including virtual network services. Moreover, as cloud servers expand, many providers want to ensure user quality support at future instants in time, e.g., for special events, sporting venues, conference meetings, etc. In the light of above, this dissertation presents one of the first studies on advance reservation of virtual network services. First, the fixed virtual overlay network scheduling problem is addressed as a special case of the more generalized virtual network scheduling problem and a related optimization presented. Next, the complete virtual network scheduling problem is studied and a range of heuristic and meta-heuristic solutions are proposed. Finally, some novel flexible advance reservation models are developed to improve service setup and network resource utilization. The performance of these various solutions is evaluated using various methodologies (discrete event simulation and optimization tools) and comparisons made with some existing strategies.

Grid computing provides the necessary infrastructure to pool together diverse and distributed resources interconnected by networks to provide a unified virtual computing resource view to the user. One of the important responsibilities of the grid software is resource management and techniques to allow the user to make optimal use of the resources for executing applications. In addition to the goals of minimizing job completion time and achieving good throughput there are other minimum requirements such as minimum memory and cpu requirements, choice of operating system, fine grained file access permissions etc. Currently such requirements are being fulfilled by resource brokers, which act as mediating agents between users and resource owners. In this thesis we approach the resource brokering architectural issue in a different manner. Instead of a monolithic broker, which performs all the superscheduling functions we propose a Modular Framework based Architecture for Task Initiation and Scheduling (MFATIC) based on the three main stages in the superscheduling process. There are three major goals of this research. The first aim is to develop a decoupled architectural model that not only provides a clear distinction in the responsibilities of each of the components but also provides the user the flexibility to replace one component with another functionally equivalent component. Secondly each of these components should be configurable and extensible to be able to accommodate user requirements. Finally, the design should enable the user to plug in modules within components of different deployments of the resource broker and thus promoting software reuse. / Master of Science

Next generation data centres are expected to support both advance resource reservation and on-demand access, but the system performance for such a computing environment has not been well-investigated. A reservation request is characterized by a start time, duration, and resource requirement. Discrete event simulation is used to study the performance characteristics of reservation systems. The basic strategy is to accept a request if resources are available and reject the request otherwise. The performance metrics considered are resource utilization and blocking probability. Results showing the impact of input parameters on these performance metrics are presented. It is found that the resource utilization is quite low. Two strategies that can be used to improve the performance for advance reservation are evaluated. The first strategy allows the start time to be delayed up to some maximum value, while the second allows the possibility of non-uniform resource allocation over the duration of the reservation. Simulation results showing the performance improvements of these two strategies are presented. Resources not used by advance reservation are used to support on-demand access. The performance metrics of interest is the mean response time. Simulation results showing the impact of resource availability and its variation over time on the mean response time are presented. These results provide valuable insights into the performance of systems with time-varying processing capacity. They can also be used to develop guidelines for the non-uniform resource allocation strategy for advance reservation in case the reserved resources are used for interactive access.

Next generation data centres are expected to support both advance resource reservation and on-demand access, but the system performance for such a computing environment has not been well-investigated. A reservation request is characterized by a start time, duration, and resource requirement. Discrete event simulation is used to study the performance characteristics of reservation systems. The basic strategy is to accept a request if resources are available and reject the request otherwise. The performance metrics considered are resource utilization and blocking probability. Results showing the impact of input parameters on these performance metrics are presented. It is found that the resource utilization is quite low. Two strategies that can be used to improve the performance for advance reservation are evaluated. The first strategy allows the start time to be delayed up to some maximum value, while the second allows the possibility of non-uniform resource allocation over the duration of the reservation. Simulation results showing the performance improvements of these two strategies are presented. Resources not used by advance reservation are used to support on-demand access. The performance metrics of interest is the mean response time. Simulation results showing the impact of resource availability and its variation over time on the mean response time are presented. These results provide valuable insights into the performance of systems with time-varying processing capacity. They can also be used to develop guidelines for the non-uniform resource allocation strategy for advance reservation in case the reserved resources are used for interactive access.