Dynamic Resource Scheduling Based On Stochastic Optimization In Cloud Environments

In recent years, due to the impetus of the governments, enterprises, market demands and so on, Cloud Computing and its related technologies (e.g. mobile cloud computing, cloud computing based on software-defined networking (SDN), etc.) have a rapid development. At the same time, the amount of cloud com-puting infrastructures and mobile devices also has an explosive growth. How to reasonably and efficiently utilize the resources of cloud computing and mo-bile devices to achieve a cost-efficient and energy-efficient resource provisioning (or service provisioning for service provider) with Quality of Service (QoS) guar-anteeing is an inevitable and practical problem. In this dissertation, based on the related stochastic optimization theories, the framework of the resource man-agement, QoS-guaranteed dynamic resource scheduling, uplink data transmission scheduling with energy-efficiency in mobile terminals, and so on are studied re-spectively. The detail of our main work is as follows:1) Firstly, for the networking systems based on cloud and SDN, in order to provide an integrated resource slice with computing resources, storage resources and network resource (OpenFlow-enabled networks) for the users, and achieve in-telligent resource scheduling and allocating, a layered and converge resource man-agement framework, software stack and a generic process for resource scheduling optimization with self-sensing and auto-scaling capacity are proposed. In addi-tion, in order to simplify the management and utilization of the network resource, a modularization for the forwarding/routing policy service is adopted and some common forwarding/routing policies are designed and implemented specially. Fi-nally, experiments on a prototype system are carried out to verify the efficiency of the modularization for the forwarding/routing policy service.2) The dynamic resource allocation with the cost-efficiency and QoS guaran- teeing in IaaS cloud platform need consider the randomness of the task request, the delay of starting up a cloud instance, etc. Therefore, in this dissertation, the dynamic instance provisioning problem for IaaS clouds is formulated as minimiz-ing the number of active instances subject to a QoS requirement in terms of the desired overload probability, and an online overload probability estimation model based on the large deviation principle is proposed. In order to adapt to the needs of different application scenarios, two basic purchasing schemes (All On-demand Scheme and Combining Scheme) are employed in this paper according to the types of instance and the pricing model of Amazon EC2. For the Combining Scheme, a Reserved Instance Provisioning strategy based on AR Model is proposed in order to further reduce the cost. The experiments were carried out to verify the per-formance of the proposed strategy using two real workload traces, and the results illustrated that the proposed strategy could adaptively provide instances for dy-namic computing demands with a good trade-off between the cost saving and the QoS requirement.3) Real-time video transcoding for a large MPEG-DASH system often re-quires a lot of computing resources and cloud computing is a perfect candidate. Therefore, the problem of dynamic resource configuration/reconfigration for video transcoding system with the aim of cost saving and QoS guaranteeing in cloud environment was studied in this dissertation. By defining the transcoding jit-ter probability as a metric of QoS, a QoS-guaranteeing dynamic transcoding re-source configuration algorithm is proposed based on the large deviation principle, which can proactively estimate the transcoding jitter probability and adjust the transcoding resources according to the result of the comparison between the es-timated transcoding jitter probability and the pre-negotiated QoS value. The experiments were carried out to verify the performance of the proposed strategy on our cloud platform based on OpenStack, and the results illustrated that the proposed algorithm can achieve a good cost saving and the QoS guaranteeing.4) Finally, in order to achieve an energy-efficient uplink data transmission with satisfying the delay constraint for the mobile terminal in the mobile cloud computing system, this dissertation formulates the connection management and packet scheduling for uplink data transmission as a Constrained Markov Decision Process (CMDP), which considers the randomness of the data arrival and the channel state. In addition, according to the idea of offloading, to further reduce the energy consumption for the mobile terminals, a system architecture is de-signed to offload the computation of solving CMDP to the cloud data center. In some real environments, the transition probability of CMDP may not be known or obtained, thus Q-learning is used in this dissertation to solve the CMDP. Fi-nally, the simulation is implemented in MATLAB to verify the performance of the proposed algorithm.