Distributed Regional Weighted Multidimensional Scaling for Network Coordinate System

Abstract

Network Coordinate System embeds all nodes into a specific metric space and estimate the latency between any node-pair through computing the corresponding metric distance with low-cost. But the metric space is usually uniform while the Internet is heterogeneous since different autonomous systems (AS) always have different routing strategies. To make full use of the clustering characteristic for improving the latency estimating accuracy without any prior knowledge, this paper judges the reference nodes are in the same AS with the target node or not through the metric distance between the target node and the source, and weights them by the Gaussian kernel function. Then the source node can computes a new coordinate to estimates
the latency between the target node and itself more precisely. The experiments show this mechanism can reduce the Stress criterion of the entire system by 19.38% without any additional communication overhead, its computation cost is also very small at the same time.

DOI : http://dx.doi.org/10.11591/telkomnika.v12i2.3817