Multi-manifold model of the Internet delay space

The network coordinate systems (NCSes) can assist the network applications to improve their performance by choosing preferred severs or constructing optimal overlay networks. However, the current NCSes cannot predict the end-to-end delay accurately, because they have neglected the inherent characters of the Internet delay space. By analyzing typical datasets, the Internet delay space was found to have obvious multi-manifold property. Thus, a multi-manifold model of the Internet delay space and a Principal Component Analysis (PCA) based Multi-manifold Partition algorithm of the Internet Delay Space (MPIDS-PCA) were proposed. Beginning with choosing landmarks randomly, MPIDS-PCA partitions the whole dataset into several sub datasets with low dimensionality by iterations. After the partition, the whole delay dataset and sub datasets are embedded into a hierarchical coordinate system Vivaldi-M. The experimental results show that MPIDS-PCA can gurantee the low dimensionality of sub datasets and Vivaldi-M can achieve better prediction accuracy.