Research On The Computational Model And Application Of Network Coordinate

Large scale Internet application is a hot topic in today's Internet research. Obtain-ing the distance accurately between the Internet hosts and doing locality-aware Internetpath selection is an important factor in improving the performance of such application.Network Coordinate (NC) System is a manner for Internet distance prediction withhigh scalability and low measurement overhead.The key problem of NC related research is the improvement of the predictionaccuracy. This thesis focuses on the following two important problems: sub-optimalrouting and node churn. Extensive studies are done for the computational model, algo-rithm, architecture and applications of NC systems and solutions with high predictionaccuracy are proposed. The main contributions and innovations are as follows.First, for the sub-optimal routing problem, extensive analyses have been donefor the matrix factorization based NC computational model. By introducing weightedmodel, a decentralized NC system called Phoenix is proposed. According to our ex-periments, Phoenix can reduce the relative error of distance prediction greatly fromexisting NC systems.Second, to guarantee the compatibility with the existing Euclidean distance basedNC system, a decentralized hierarchical NC computational model is proposed. A de-centralized and hierarchical NC system called Pharos is designed and implementedby using this model. The experimental results show that Pharos greatly outperformsVivaldi in Internet distance prediction. Our extensive evaluation results also demon-strate that Pharos can significantly improve the performance from Vivaldi in distributedInternet applications, such as closest neighbor selection and overlay multicast.Third, for the node churn problem, a decentralized NC computational modelwhich is robust to node churn is proposed. A decentralized NC system called Mythis designed and implemented by using this model. Our experimental results show thatMyth is a cost-e?ective way to achieve higher prediction accuracy, it will not only improve the prediction accuracy but also save the communication overhead.Last, A NC based application layer anycast system called Proxima is proposed.We have implemented and deployed Proxima in Internet. By using Pharos NC system,Proxima can achieve low latency, high bandwidth server selection from the anycastgroup. Proxima is a promising solution to achieve high performance global anycastservice.