A Comparison of Large-S ale Overlay Management Te hniques

Abstra t In this paper we present Kudos, a novel hierar hi al, topology aware overlay onstru tion algorithm. Kudos is an extension of Narada, an existing topology aware overlay; by adding hierar hy, we have signi antly in reased the s alability of Kudos while maintaining the performan e advantage of topology awareness. Additionally, we provide the rst detailed omparison between topology aware and topology agnosti overlay onstru tion algorithms, by performing detailed simulations of Kudos, CAN [1℄, Chord [2℄, and a \random power law" overlay. We show that all overlay topologies are fa ed with a fundamental tradeo between relative delay penalty (RDP) and link stress, and that by hanging node out-degree, di erent points in this tradeo an be sele ted. We demonstrate that Kudos signi ant outperforms all onsidered topology agnosti algorithms, both in terms of laten y and bandwidth, although it annot s ale to the same degree be ause of the osts in urred by topology aware overlay maintenan e.