Research On Resource Scheduling Method Of Data Center For Intelligent Video Surveillance Network

The intelligent video surveillance network, which consists of massive intelligent surveillance cameras, has already covered most of the main roads and buildings in the city. The network can provide 7x24 hours intelligent video surveillance services for the end users. These services are usually deployed onto the video servers,and each video server manages several surveillance cameras. However, the servers don’t share the information, and the limitation of the server’s resources obstructs the improvement of the services. The video data center, which manages large-scale server cluster, can avoid the drawbacks of the video servers. The inter-connected video data centers can both share the information and utilize the resources elastically in the cluster or among the clusters. Therefore, more and more intelligent surveillance services are migrated to the video data center. For this reason, it is very urgent to efficiently assign the resources,and improve the service quality on the video data centers.The video data center gains video streams from the intelligent video surveillance network, forwards the video streams, and analyzes the video streams to provide the video surveillance services. In this thesis, we mainly focus on the research of the resource scheduling problem of the video data center,which includes three aspects:video stream uploading, forwarding and processing. Furthermore, the factors such as service delay,resource prices,resource availability and scalability must be considered by the resource scheduling approach. Then, 1) for the video stream uploading, we deal with the resource scalability problem. 2) For the video stream forwarding, we research the forwarding route selection problem under the multiple constraints. 3) For the video stream processing, we focus on the server allocation and runtime workload balance problems. The main contributions of the thesis are as follows:(1) An upload stream scheduling approach based on the "selfish+shared" resource slot strategy. When video data center answers the live video uploading requests, it needs to provide the resource’s availability and scalability. We first formulate the service capability of a server as a number of the resource slots. Furthermore, we give a comprehensive consideration on the server’s cost, and the network cost on both uploading and forwarding stages to build up the models on the resource availability and scalability. Based on the models, we minimize the total resource costs by utilizing the selfish and the shared scheduling strategy to adjust the network and server cost. The experimental results indicate that our approach can adapt to different scale of the end users and different resource pricing mechanisms, and efficiently cut the total resource costs to provide resource scalability.(2) An online self-organized video stream forwarding approach based on the primal-dual optimization. For the route selection problem of the inter-connected video data centers, we give a consideration on the factors of the forwarding delay, the resource cost, and the resource availability as well as scalability to comprehensively evaluate the resource provisioning cost. Then, the resource utilization model and the video forwarding model of the video data centers are built up in the thesis. Based on these models, we aim at optimizing the resource provisioning cost within an acceptable time,and gain an approximate optimization with a specified competitive ratio. We sequentially deal with 1) the facility location problem, and 2) the shortest path problem to reduce the complexity of the problem. For the facility location problem, we utilize the primal-dual optimization method to gain the media server of the proper data center with less resource cost and forwarding delay, and the chosen media server can directly forward the live video stream to the end user. For the shortest path problem, we further employ a greedy algorithm to find the forwarding path with the minimum resource provisioning cost. The experimental results indicate that our online approach can achieve the equilibrium among multiple factors, and efficiently cut down the resource provisioning cost.(3) A multi-type server allocation approach based on the critical path of the task workflow. The user employs the workflow of the video processing task to describe the sub-tasks of the task and their relationships, and submits the workflow to the video data center. Our approach utilizes the process capability and the server cost to characterize each sub-task, and calculates the task finishing delay by finding the critical path of the workflow which consists of several critical sub-tasks. The resource scheduling approach finds the critical path of the workflow at first, and assigns different types of servers on the critical path with optimal capability-cost ratios to meet the user requirement of both server cost and task finishing delay. Then, the rest of the sub-tasks gain the minimum number of servers. The experimental results indicate that our approach can efficiently achieve the balance between the server cost and finishing delay, when it meets the user requirements.(4) A server scheduling approach for the service delay guarantee. At runtime, the video data center needs to support that the actual finishing delay of a sub-task is not later than the required sub-task finishing delay. However, the number of the receiving workload is not fixed each time, and the cumulative workload on the server cluster can lead to longer queuing delay. To deal with this problem, we first formulate the queuing workload of the server cluster as a queuing system, and then evaluate the workload skew without depending on the workload predictions. Based on this model, we utilize the Lyapunov optimization method to adaptively add and release the servers according to the changes of the input workload,and guarantee the sub-task’s service delay. The experimental results show that our approach not only pays the minimum cost to support the sub-task’s service delay, but also maintains the stability of the workload queuing system.Finally,we design and implement a video data center prototype system which provides the intelligent traffic surveillance services. The prototype system consists of three layers, i.e., the resource layer, the platform layer and the application layer. They correspond to the management of the physical resources, the virtualization resources,and the surveillance services. All the proposed approaches in the thesis are deployed onto the prototype system to schedule the virtualization resources. The experimental results show the effectiveness of our approaches.