Design And Implementation Of Web Service QoS Prediction System For Multiple Attributes

With the rapid development of Web service, there is a serious need to select the best one from various services for users. Quality-of-Service (QoS) can provide important insights for service recommendation and selection. Based on distributed network coordinate system, a prediction system for multiple QoS attributes-response time, bandwidth and reliability, is designed and implemented, while the corresponding prediction approaches are proposed.The new prediction approaches for these three attributes are designed after studying the characteristic of QoS attributes and analyzing the advantages and disadvantages of the existing methods. Firstly, as most of existing prediction methods for response time suffer from the sparsity of historical QoS data, which limits the accuracy of prediction result, distributed network positioning technology is introduced in our method. Services and users are mapped into a multi-dimensional coordinate system and located by a large number of measurement nodes. Secondly, for measurement of available bandwidth, the existing methods cannot applied directly in this scenario, so an ICMP-based approach without service-side application is proposed. Thirdly, to predict reliability, clustering is used to find similar users according to their coordinates, which addresses the problem of sparsity of QoS data and takes advantage of the network coordinate system. Moreover, this paper completes the requirement analysis, design and implementation of the prediction system on basis of the methods above. The prediction system contains many measurement nodes and some other functional components. The measurements nodes measure QoS periodically and position services, and those functional components collaborate on the task when there is an online prediction request from user. Finally, the system is deployed and tested on PlanetLab and the real-world QoS data collected on the PlanetLab are utilized to conducted extensive experiments. The experimental results show that the feasibility and effectiveness of our methods can meet the actual demands.