Research On The Efficient Resource Scheduling In Cloud Data Centers

With the rapid development of information industry and the widely spreading of Internet technology, data center plays an important role in supporting the informational service. More and more companies transfer their applications to large scaled data center which hosting numerous computing resources. Thus, the scale of the data center is undergoing a rapid expansion. The booming trend also leads a lot of problems, in which the resources scheduling is one of the most crucial issues to foundation of data centers. The contradiction between inefficient costly resource provision mechanism and the sharply rised resource demand challenges the sustainable growth of data centers. Therefore, it is necessary to seek new ways of building next-generation data center under the urgent need of improving the infrastructure utilization.Cloud computing tends to leverage its newly resource provision pattern helps data center overcoming this weakness. The goal of massively scalable cloud data centers is to make applications reside in where computational resources can be dynamically provisioned and shared to achieve significant economies of scale. How to design the efficient resource scheduling mechanism of cloud data center is a considerable problem. Faced the new challenge of resource sharing and dynamic management involved by virtualization, new researches will be raised to meet the need of on-demand, elastic and scalability in cloud data centers. Therefore, this dissertation focuses on the key techniques of improving the efficiency of cloud data center resource scheduling. Based on analysis of current techniques, we make an in-depth researching on the potential methods on the enhancement of resource utility. The main contributions of our work include:(1) We proposed a multidimensional coordination based virtual machine (VM) scheduling mechanism. It aims at improving the resource utilization of data center by balanced usage of multiple resources, as multi-tenant technology in cloud computing allows heterogeneous virtual machine share a virtualized data center. The problem is described as a vector packing model with multi-dimensional coordination. We carried out a group genetic based multi-dimensional coordination scheduling algorithm, which leverage the advantage of group characteristic. To guide the solution searching, a fuzzy logic based multidiemsional fitness function is raised. As well, innovative optimization of key operators is put forward to improve the solution quality. The experiment results proved that our algorithm would efficiently reduce the imbalance in multiple resources and finally increased the utilization of the infrastructure resources in cloud data centers.(2) In order to efficient use of the network resource in cloud data centers, we proposed a communication-aware multi-tier application placement problem with conflict-avoidance. The network resource in data centers is often hierarchical organized and limited, which directly affects the entire performance and the scalability of data centers. The cost of network resource is closely related to two parameters namely traffic patterns among VMs and the topology of the data center networks. For the purpose of increasing scalability of data center networks, we prefer mapping much heavier traffic onto low-cost links to relax the communication load. The C2VMPP is modeled as a variant of Quadratic Assignment Problem, and a dual-stage optimization placement algorithm with the consideration of multi-tier application characteristics is introduced to obtain an approximate solution of C2VMPP. The evaluation results of DOPA verified our motivation for supporting higher scalability by optimizing the network utilization and balancing the total communication load.(3) In order to providing both the flexibility and time-efficiency of large-scale resource sharing in data centers, we focus on the scheduling mechanism of the parallel sequence oriented VM migration problem. The high consolidation level of infrastructure in cloud data centers results in more sequence constraints when the hypervisor tries to migrate a great number of VMs at the same time. Moreover, the delay caused by live migration exacerbates the challenge of large-scale VMs parallel scheduling. Throughout this paper, we first modeled the scenario of parallel sequence oriented VM migration problem as a flow-based graph, and then described an effective method for total migration delay evaluating from the perspective of VMM. Based on the principal of "min-max", we proposed a Scheduling Algorithm of Adaptive Migration (SAAM) to solve the parallel sequence oriented VM migration problem, which works by reducing the generating speed of dirty pages and/or adjusting bandwidth for migration activity. The simulation results indicate that not only our SAAM algorithm would decrease the migration delay for a single VM migration operation, but also contribute to reducing the total migration delay of parallel VM migration with sequence constraints according to the flow-based modeling. As a conclusion, SAAM algorithm is helpful to optimize the delay of live migration for both single and parallel VM migration scenarios.