Research On Efficient Publish/Subscribe Systems In Virtual Computing Environment

With the rapid development of the information technology, Internet is becoming the basic information infrastructure of modern society. Internet-based Virtual Computing Environment (iVCE) is a novel network computing platform which fits the features of the Internet. iVCE is based on the autonomisation of the internent rescourses and focuses on resource sharing and collaborative working over the open network infrastructure, with the key mechanism of on-demand aggregation and autonomic collaboration. The aggregation of distributed, heterogeneous and dynamic internet resources requires a loosely coupled interaction mechanism. The publish/subscribe systems is an important way to aggregate resources with its loosely coupled features in time, space and control flow.iVCE uses content-based publish/subscribe systems to aggregate resources. The content-based publish/subscribe systems can be categorized as two types: broker-based and broker-less, each of which is applicable for applications with different network conditions in iVCE. The filter-based system is the most widely deployed broker-based system. The broker-less systems can be further categorized into name-based and value-based, each of which is applicable for applications that require different expressive ability in iVCE. This thesis focuses on the efficient realization of different publish/subscribe systems used in iVCE.Firstly, the thesis proposes a generic model of event matching for name-based systems, and proposes popularity-based event matching method based on the generic model. Current event matching methods all consider the matching problem from the event aspect, and can be categorized into event-oriented methods; and considering the matching problem from the subscription aspect shows another design possibility for matching problem, and can be categorized into subscription-oriented methods. The generic model is a genelization of these two methods. PEM (Popularity-based Event Matching) uses the attribute set to distribute the load and uses the popularity information of attributes to control the bandwidth cost. The correctness of PEM is theoretically proved. The theoretical analysis also shows that PEM can greatly reduce the bandwidth cost during event publishing while only induces a few more subscriptions stored in the system compared to current event-oriented methods. For applications with small number of attributes, PEM uses"keys regeneration"to further distribute the load.Secondly, the thesis proposes the indexing structure based on attribute grouping for value-based systems. Such indexing structure is the generalization and improvement of current indexing structures: S-MD (a Single Multiple-Dimension indexing structure) and M-1D (Multiple Single-Dimension indexing structures). In attribute grouping, the whole set of attributes are divided into several groups, each of which contains some attributes. The subscription chooses one group and is indexed based on the attributes in the group. The event chooses all groups and for each group, the event is indexed based on the attributes in the group. The indexing structure and the method to construct overlays are proposed based on attribute grouping. The results from the experiments show that attribute grouping can achieve better balance between the system load and the bandwidth cost.Thirdly, the thesis proposes a novel subscription merging mechanism to reduce the subscription maintenance cost in filter-based systems. Compared to current subscription merging mechanisms, the main difference is that it utilizes the similarity of subscriptions. The judgement of similarity of subscriptions is based on the size and the location of constraint range of each attribute in the subscription. The proposed mechanism uses a two-phases grouping algorithm to group the similar subscriptions, and uses the similar idea of hash storage in file system to link and merge subscriptions in the same group. It can achieve low time complexity for it needs to scan the subscriptions only one time, and it can flexibly control the merging granularity according to the load condition. The results from the experiments show that the proposed method only induces a little more bandwidth cost for event forwarding.Finally, the thesis proposes two subscription tree reconstruction methods to deal with the problem caused by the movement of subscribers in the filter-based systems: STRN (Subscription Tree Reconstruction from New borker) and STRO (Subscription Tree Reconstruction from Original tree). STRN constructs the subscription tree from the new broker and utilizes the structure of the original subscription tree to reduce the handoff cost. STRO constructs the subscription tree from the original subscription tree, and converts the structure of the original subscription tree to reduce the delay of event forwarding. The results from the experiments show that STRO is convenient for the environment where subscribers move frequently for its low handoff cost, and STRN is convenient for the environment where subscribers move less frequently for its low delay for event forwarding.