Location-aware Web Service Performance Prediction

Web service performance is among the key factors concerned by users in Web serviceselection. Since the performance of most candidate Web services is unknown to a service user,performance prediction of services is therefore required. How to improve the accuracy ofservice performance prediction has recently become a hot research topic. Several recent workshave considered the influence of locations of service users or Web services on Web serviceperformance, and thereby developed a couple of location-aware Web service performanceprediction methods. However, the relationships among Web service performance, userlocations and service locations are rarely investigated in practice.This work is an empirical study on evaluating the correlation among Web serviceperformance, user locations and service locations. We use geographical locations and networklocations respectively and conduct a set of experiments on a real large-scale Web servicedataset. The geographic location based experiments map IP addresses of both users andservice to latitudes and longitudes, and calculate the geographic distance between them. Itevaluates the impact of geographic locations on the service performance, and investigates therelation between the geographic proximity among users (services) and the similarity amongusers (services). The network location based experiments define the users (services) in thesame country (AS) as neighbors. It evaluates the impact of network locations on the serviceperformance, and investigates the relation between the network proximity among users(services) and the similarity among users (services).Both of the above two location-based experimental results show:(1) The users close to aWeb service usually perceive much better performance than remote users of the Web service;(2) Nearby users have a high probability to perceive similar performance on the same set ofWeb services;(3) The user (services) close to each other have higher similarity degree thanthat of the user (services) far from each other;(4) If two users (services) have similar QoSexperiences, then it is highly probable that the two users (services) are close to each other, andvice versa. These results confirm the strong correlation among Web service performance, userlocations and service locations, and provide important evidences for the effectiveness oflocation-aware Web service performance prediction. Based on the above experimental results,we develop several geographic location based and network location based Web serviceperformance prediction methods. Compares with existing classic Web service performanceprediction methods, we show that the proposed location-based web service performanceprediction methods will improve the prediction accuracy.