Research On Resource Organization And Scheduling Techniques In Video Grid

With the advantages of the high resource sharing and extensive service collaboration, grid technology provides an effective way to support large-scale integrated network video services. In recent years, researches on video services in grid have become a hot topic in grid technology. Video grid is a kind of service grid. It can efficiently organize and manage video resources, flexibly generate and integrate various video services, and efficiently schedule all kinds of grid resources according to the requirements of video services. In this way, it can support video services as a uniform platform. For managing and using video resources efficiently, the video resource organization mechanism and scheduling algorithm become the key problem in video grid. They are very complicated because of the wide variety of video services, the great requirements for video resources, and the complex QoS requirements.In this thesis, we mainly study some key problems in resource organization and scheduling. In the organization model of video resources, we concentrate on the resource organization mechanism surpporting multi-keywords query and range query. Based on the organization model of video resources, we deeply study the scheduling algorithm of video processing services and streaming services. Moreover, based on the resource organization mechanism and scheduling algorithm, we develop a prototype system for face recognization in video sequence, which indicates the application value of our research works.This thesis has mainly made some contributions as follows:(1) This thesis proposes a hierarchy resource organization mechanism based on DHT and extend clusters to meet the requirements of multi-keywords query and range query. We adopt a two-layer structure. Because single keyword locating in DHT has time complexity of Log₂N, and the classified information of video resources is stable, we adopt advanced DHT to organize video resources in the upper layer. In the lower layer, we adopt extend clusters to organize video resources flexibly in the resource class according to the detailed description information. Based on this structure, we can achieve multi-keywords query and range query for video resources quickly and efficiently.(2) To meet the real-time requirements of large-scale network video processing services, this thesis proposes a heuristic resource scheduleing algorithm based on grid node clustering. We adopt cluster method to group the grid nodes to one cluster to meet application's scheduling requirements, which reduces the node searching space and scheduling scope. In node class, we schedule nodes by bandwidth-first method, and achieve a balanced distribution of tasks assigning through equational adjustments method to make video content processing services completed in the shortest time.(3) This thesis proposes a two-level switching algorithm based on probability forcasting to meet the high stability requirement of video streaming services. First, according to the difference of performance and stability among terminals, as well as the invalidation/congestion irrelativeness among terminals and network paths, we find the switched node sets with the whole switching invalidation probability method. Second, we adopt probe switching algorithm to reduce the cost of the failure switching between video servers with low cost and low stability. Third, we make full use of transmitted video streaming information during switching process by two-level switching, so that the stability of video streaming is guaranteed with low cost.(4) This thesis proposes a local limited deforming distributed elastic graph matching algorithm based on face content, in order to meet the character of elastic graph matching face recognition algorithm, such as large amount of data and high time complexity. In distributed face recognition algorithm, we partition face by multi-level region partitionning method based on geometrical properties of face, and rapidly recognize face regions detailedly by local limited deforming elastic graph matching algorithm. Furthermore, because every region can independently perfom matching process, we schedule these matching processes to grid nodes for parallel matching, in order to reduce the recognition time.