Research On Some Key Technologies Of Service-Overlay Network

With the development of service-driven economy and the mature of the service industry chain, the corresponding research areas represented by service science and service computing have developed fast. Web service technologies and SOA (Service Oriented Architecture) are widely adopted by IT industry. Large numbers of online services have emerged. Web services are used with scalability, evolvability, dynamics and flexible resource demand. To overcome the challenges from the wide usage the large-scale web service flexibly and efficiently, complexity of the service systems and increasing cost of implementing and maintaining, we have to research on the areas with networked and complex views.Based on the complex network theory and distributed mind, we research on several key technologies on how to enable service discovery, selection, manage and implement services more efficiently and conveniently to manage and organize the large-scale web service in the service overlay network. To ensure the widely usage and efficiently processing of web services, we research on how to select web service efficiently, how to deliver the proper web service and how to implement web service flexibly under the flexible and networked web service delivery platform based on cloud. And we research on the corresponding key technologies include service clustering, QoS (Quality of Service) prediction of service, service routing and host load prediction in cloud clusters. Service users firstly select proper service with distinguishing the function and QoS of services, and then receive the service by efficient service routing, and at last implement the service on the proper host by predicting the host load for service load balance and service migration.The major contribution of this thesis is as following:1) To deal with the low efficiency of service selection among the large number of services, we propose TopLPA (Top Label Propagation Algorithm) based on constructing service similarity network. That algorithm extracts the features from the service description and combines them with semantic feature to ensure the reliability and productivity in service similarity computation. TopLPA distinguished the neighbors while refreshing the class of the nodes and its accuracy is better than LPA (Lable Propagation Algorithm)5.17%averagely. And to improve the efficiency of the algorithm while new service arrived, we modify TopLPA as TopLPA-Online, which can reduce the time21%. That feature makes TopLPA-Online is much more practical.2) To deal with the spare problem in service QoS prediction for large-scale services in service overlay network, we propose the link prediction algorithms to find implicit neighbors to evaluate accuracy and success rate. We first construct service user similarity network and service similarity network, and then modify the link prediction algorithms to find the implicit neighbors. While we modify the link prediction, we consider the location proximity of the service users and the "weak tie" affect. Simulation results show that HUWAA (Hybrid User-Location-Aware Prediction based on Weighted Adamic-Adar) and HUWRA (Hybrid User-Location-Aware Prediction based on Weighted Resource Allocation) are both better than UPCC (User-based Pearson Correlation Coefficient), IPCC (Item-based Pearson Correlation Coefficient) and WSRec (Web Service Recommendation). And HUWAA performs7.59%,5.77%,48.1%better and IPCC in MAE (Mean Absolute Error), RMSE (Root Mean Square Error), and FR (Failure Rate) respectively. HUWRA performs10%,8%,48.8%better and IPCC in MAE, RMSE, and FR respectively.3) To deal with the bottleneck of performance and point of failure introduced by centralized management of service discovery in service overlay network, we propose the distributed routing algorithm TISRQ. That algorithm constructs the index based on the term set and considers the term popularity in service users’requests. It decreases the number of indexing and also evaluates the accuracy. Meanwhile, it extends MAAN to support QoS value range query. Simulartion results show TISRQ and TSM both can reduce75%indexing compared to single term indexing. And TISRQ performs6.09%better than TSM in accuracy at most. To enable the flexible use of service, we propose the corresponding service evaluation method which allows SLA (Service Level Agreement) violations. That method enables service users define dynamic SLA and form service evaluation preiodically.4) To predict host load accurately in cloud system for resource management, we propose the features for prediction based on classification and conduct the experiments for practice. Firstly, we transform the prediction problem into classification problem based on the existing ESP (Exponentially Segmented Pattern). And then we propose our new feature with existed features to describe the host load more efficiently and accurately. Experiments on the real world dataset invalidate the efficiency of the new proposed feature, which is in the most effective combinations of features and it performs better than the existed feature the finess index. And the results also show that SVM method can achieve nearly the same performance as the Bayes methods and the performance is about50%higher in SR (Successful Rate) and17%better in the MSE (Mean Square Error) compared to the existed methods.