Two-layer network coordinate system for Internet distance prediction

Abstract

Network coordinate (NC) system is an efficient and scalable system for Internet distance prediction. In this paper, we propose 3 two-layer NC systems HNPS, HBBS and HIDES derived from NPS, BBS and IDES, respectively. We use hierarchical idea in both Euclidean coordinate systems (NPS and BBS) and matrix factorization model (IDES). And to the best of our knowledge, we first introduce the thought of hierarchy into matrix factorization model. By comparative simulation, we evaluate the performance of HNPS, HBBS, HIDES, NPS, BBS and IDES systems in both latency data of King and Planetlab. Comparing with NPS, BBS and IDES, we find that HNPS, HBBS and HIDES have better distance prediction, respectively. By in-depth analysis, we reveal that the essential reason why two-layer NC models outperform the ones of one-layer is that two-layer NC systems reduce prediction errors of short link distance. In addition, we compare the performance of HNPS, HBBS and HIDES with that of NPS, BBS and IDES, respectively in overlay multicast application. The experimental results show that two-layer NC systems achieve higher accuracy in Internet distance prediction with lower cost of overlay multicast tree construction. Therefore, the thought of hierarchy is not only suitable for Euclidean distance based NC systems, but also suitable for matrix factorization based NC systems, which fully substantiates that two-layer NC system is a general model for accurate estimation of network distance.