bar-gossip-apr-2006 - BAR Gossip Harry C. Li, Allen...

Info iconThis preview shows pages 1–2. Sign up to view the full content.

View Full Document Right Arrow Icon
BAR Gossip Harry C. Li, Allen Clement, Edmund L. Wong, Jeff Napper, Indrajit Roy, Lorenzo Alvisi, Michael Dahlin Dept. of Computer Sciences, The University of Texas at Austin We present the first peer-to-peer data streaming ap- plication that guarantees predictable throughput and low latency in the BAR model, in which non-altruistic nodes can behave in a self-serving (rational) or even arbitrarily malicious (Byzantine) way. At the core of our solution is a BAR-tolerant version of gossip, a well-known technique for scalable and reliable data dissemination. In traditional gossip, data dissemina- tion is performed with randomly selected partners; such non-determinism, we show, offers an excellent opening to rational nodes bent on gaming the sys- tem. BAR-tolerant gossip instead relies on a veri- fiable pseudo-random partner selection mechanism that eliminates non-determinism while maintaining the unpredictability and rapid convergence of tradi- tional gossip. Our initial experience indicates that BAR Gossip is robust against up to 20% of nodes exhibiting Byzantine behavior and even against up to 40% of nodes colluding together against the re- maining nodes. In either case, our BAR-tolerant video streaming application provides over 95% con- vergence for broadcast updates. 1 Introduction Streaming media is an increasingly useful application at several scales of deployment. For example, the 2006 NCAA tournament had peak participation of 150k+ users for their live streaming services. At smaller scales, such as academic conferences like OSDI or artistic events like Austin’s SXSW, venues draw audiences of dozens to hundreds of people. At all these scales, a peer-to-peer (p2p) streaming so- lution appears to be an intriguing alternative to tradi- tional methods. One advantage of p2p systems is their potential to be highly robust, scalable, and adaptive. For example, a p2p architecture could, in principle, absorb the impact of an unexpected flash crowd. Furthermore, large-scale content providers may adopt p2p-based so- lutions to shift costs (like bandwidth) to clients, and small-scale providers might find it simpler to use a self- organizing p2p network instead of maintaining a dedi- cated server. Realizing the promises of p2p in streaming services is a non-trivial task. First, the service needs to guaran- tee highly reliable, stable, and timely throughput of mes- sages despite the presence of faulty, misconfigured, or even malicious peers. Second, the service must be ro- bust against selfish users, who try to catch a free ride by receiving streams without contributing their fair share to others [16]. The freerider phenomenon in p2p sys- tems is indeed a symptom of a broader issue: any system in which nodes are not under the control of a single ad- ministrative domain must be designed for the possibility that nodes will deviate from their specification if doing so is to their advantage. File-sharing p2p applications, like BitTorrent [10], have recognized this issue and in-
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
Image of page 2
This is the end of the preview. Sign up to access the rest of the document.

Page1 / 15

bar-gossip-apr-2006 - BAR Gossip Harry C. Li, Allen...

This preview shows document pages 1 - 2. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online