A multicast congestion control mechanism for reliable multicast

This paper evaluates a congestion control mechanism for reliable multicast applications that uses a small dynamic set of group members or representatives to provide timely and accurate feedback on behalf of congested subtrees of a multicast distribution tree Our algorithm does not need to compute round trip time RTT from all receivers to the source nor does it require knowledge of group membership or network topology Through simulations we evaluate our algorithm with and without TCP cross tra c This study demonstrates that while our algorithm makes use of bandwidth when available it responds to congestion in a timely fashion thereby limiting data drops When competing for bandwidth with TCP tra c our algorithm gracefully relinquishes bandwidth While not as aggressive as TCP it does grab a reasonable portion of the bandwidth yet keeps the overall multicast packet loss low Introduction The Internet relies on applications performing congestion control to react to network congestion and avoid congestion collapse Most applications in use on the Internet employ TCP s congestion control algorithms The increasing popularity of group communication applications such as multi party teleconferencing tools and information dissemination services motivated the development of several multicast transport protocols layered on top of IP multicast for e cient multipoint data distribution While TCP s point to point model treats multipoint data delivery as a collection of point to point ows protocols using IP multicast avoid sending duplicate data repeatedly over the same network links However multicast protocols can potentially cause considerably more harm to the network than point to point protocols To allow multicast protocols to be safely deployed on the Internet it is imperative that they incorporate mechanisms for handling congestion While many proposals have been forthcoming on the reliability aspect of multicast protocols few of them have focused on congestion control mechanisms In this paper we propose a scalable responsive congestion control mechanism for reliable multicast We focus on congestion control rather than error recovery The proposed congestion control mechanism to intended to be incorporated into current and future reliable multicast protocols Unlike point to point communication where the source reacts to feedback from a single receiver congestion control in a multicast environment requires the source react to feedback from an arbitrary number of possi bly heterogeneous receivers One approach to solving the heterogeneity problem is to nd the group s lowest bandwidth constraint and set the sender s rate to this value However this solution does not take advantage of available bandwidth and does not adapt to changing network conditions Reliable multicast protocols also face the feedback implosion problem which becomes critical as multicast group size increases Several existing reliable multicast transport protocols use probabilistic suppression to limit the amount of feedback received at the source SRM for example performs probabilistic suppression based on the round trip time RTT measured between receivers and the source The proposed multicast congestion control mechanism does not rely on RTT computation between all receivers and the source Not having to perform RTT estimation between all group members is particularly bene cial for applications such as multipoint data dissemination which employ a one to many instead of a many to many communication model Our congestion control algorithm does not require knowledge of group membership or network topology yet can react to congestion on the order of RTT to receivers in the group Our algorithm is based on the assumption that in a multicast group a small set of bottleneck links will cause the majority of the congestion problems We dynamically select a small set of group members to represent the congested multicast subtrees This small set of representatives allows us to apply the point to point congestion control model to a multicast group of arbitrary size These group representatives provide immediate feedback which suppresses feedback from other receivers thus preventing feedback implosion at the source In addition to suppression representative feedback allows the source to concentrate its congestion control e orts on the congested subtrees Based on the feedback from representatives the source adjusts the current transmission rate As new congestion appears in the tree new representatives are selected and old ones dropped from the representative set Previous work reported preliminary simulation results of the proposed algorithm s feedback control mecha nism In this paper we evaluate the proposed congestion control algorithm as a whole using a packet level network simulator In the following section we describe our model Section presents the algorithm in detail Our evaluation methodology and results are described in Section Section puts our work in perspective with related work and Section describes our plan for future work The Model Our congestion control mechanism assumes the following model including services provided by lower layer protocols and application level requirements IP Multicast Senders transmit data packets using Internet multicast There is no centralized group manage ment Hosts can join or leave the group at any time Any host can transmit to the group Single Data Source All data is sent from a single source The source is a member of the multicast group Receivers send control messages but no data Congestion control is done on a per source basis Multiple sources can be accommodated by running multiple instances of the algorithm Aggregate congestion control is subject for future work and outside the scope of this paper Unknown Group Membership Like IP multicast to support scalability we assume the set of receivers is unknown In a real implementation there needs to be some control over membership to prevent denial of service attacks Authentication is outside the scope of the congestion control mechanism Unknown Network Topology No knowledge of the underlying physical network topology is assumed Our approach is end to end routers are not relied on to provide feedback about network conditions or lter feedback requests Receiver Support Our algorithm does not require receivers to measure RTT to the source Representa tives are the primary source of feedback while other receivers only provide feedback as a back up in case of representative failure Protocol Stack Architecture When moving from a one to one reliability protocol such as TCP to a one to many protocol it becomes much more di cult to build a single protocol that can handle all the possible semantics that might be required by multipoint applications For instance higher level applications might require an explicit join protocol or might guarantee atomic delivery Others might only provide best e ort delivery to receivers Furthermore certain services may establish a minimum transmission rate constraint sources will not transmit below that rate and feedback from receivers that cannot handle the minimum transmit rate will be ignored An explicit join protocol may reject members that are unable to cope with the service s minimum transmit rate As previously mentioned the proposed mechanism focuses on congestion control an essential component of reliable multicast protocols We assume that error recovery is performed by a separate mechanism within the reliable multicast framework Algorithm Based on the assumption that only a small set of bottleneck links cause the majority of congestion problems our congestion control algorithm dynamically selects a small number of receivers or representatives to provide feedback for a subset of the group members Representatives provide immediate feedback allowing the source to adjust to the current state of the network in a timely fashion Furthermore representatives limit the amount of feedback generated by suppressing feedback from other receivers Probabilistic suppression is used as backup in regions of the multicast tree not yet covered by the current representative set Figure illustrates the concept of representatives using an arbitrary multicast tree as example Figure a shows the multicast tree with the source at the root of the tree and receivers at intermediate nodes and leaves When our algorithm starts Figure b the representative set is empty and since there is no indication of congestion all participating receivers are eligible to become representatives The source selects the receiver from which it receives feedback rst Figure c Since receiver s feedback timers are initially set to an arbitrarily large value currently second the source will probably receive feedback from the closest receivers rst When congestion is detected in the right subtree Figure d which is not yet covered by the current representative set a new representative is selected to cover the newly congested subtrees Figure e Newly congested subtrees will likely generate more feedback when congestion is rst detected since feedback control relies exclusively on probabilistic suppression until a representative is selected Receiver Receiver Receiver Receiver Receiver Receiver Receiver Receiver Source a Receiver Receiver Receiver Receiver Receiver Receiver Receiver Receiver Source