Related Work

P2Cast is related to some previous works in the context of application-level multicast systems, peer-to-peer steaming services, and available bandwidth measurement techniques. To overcome the lack of IP multicast, many application-level multicast systems have been proposed recently. e.g., [6,7,14,8,9,15,16,17]. Many of them use the delay as the single or primary metric in the tree construction, attempting to minimize the average delay among peer nodes. To satisfy the high bandwidth requirement of applications like streaming, some tree construction algorithms resort to limiting the out-degree of peer nodes [17]. In contrast, P2Cast uses the information of measured available bandwidth in the base tree construction. The available bandwidth larger than the playback rate is guaranteed over the entire base tree. Several peer-to-peer streaming systems have been developed [18,19,20,21]. Peercast [18] and CoopNet [19] are mainly designed for live media streaming. CoopNet also supports VoD service, and employs ``distributed streaming'' to obtain the content from multiple peers simultaneously. SplitStream [21] is a high-bandwidth content streaming/distribution system that is built upon Pastry [22], a generic substrate for peer-to-peer applications. There are several efforts put forth by industry, such as Allcast [23], vTrails [24], and Bluefalcon [25], that claim to provide live streaming and on-demand service. However, we cannot do a specific comparison because of the absence of published information on their on-demand services. Finally, estimating the available bandwidth efficiently is very important for the overlay construction algorithm in P2Cast since it determines clients' joining delay and the likelihood of providing continuous playback in the face of disruptions. Reference [13] claims that their tool needs less than 15 seconds to produce an estimate of the available bandwidth. We expect the bandwidth measurement in P2Cast takes less time since the granularity of bandwidth of interest in P2Cast is the video playback rate. The measurement overhead can further be reduced if the BF-delay-approx algorithm is used. Furthermore, since the available bandwidth measurement has little impact to other traffic flows [13], we believe that the concurrent bandwidth measurement toward the same requesting client will not affect the measurement accuracy significantly.