Enabling Conferencing Applications on the Internet usingan Overlay Multicast Architecture∗Yang-hua Chu, Sanjay G. Rao, Srinivasan Seshan and Hui ZhangCarnegie Mellon University{yhchu, sanjay, srini+, hzhang}@cs.cmu.eduABSTRACTIn response to the serious scalability and deployment con-cerns with IP Multicast, we and other researchers have ad-vocated an alternate architecture for supporting group com-munication applications over the Internet where all multi-cast functionality is pushed to the edge. We refer to such anarchitecture as End System Multicast. While End SystemMulticast has several potential advantages, a key concernis the performance penalty associated with such a design.While preliminary simulation results conducted in static en-vironments are promising, they have yet to consider the chal-lenging performance requirements of real world applicationsin a dynamic and heterogeneous Internet environment.In this paper, we explore how Internet environments andapplication requirements can influence End System Multi-cast design. We explore these issues in the context of audioand video conferencing: an important class of applicationswith stringent performance requirements. We conduct anextensive evaluation study of schemes for constructing over-lay networks on a wide-area test-bed of about twenty hostsdistributed around the Internet. Our results demonstratethat it is important to adapt to both latency and bandwidthwhile constructing overlays optimized for conferencing ap-plications. Further, when relatively simple techniques areincorporated into current self-organizing protocols to enabledynamic adaptation to latency and bandwidth, the perfor-mance benefits are significant. Our results indicate thatEnd System Multicast is a promising architecture for en-abling performance-demanding conferencing applications ina dynamic and heterogeneous Internet environment.∗This research was sponsored by DARPA under contractnumber F30602-99-1-0518, and by NSF under grant num-bers Career Award NCR-9624979 ANI-9730105, ITR AwardANI-0085920, and ANI-9814929. Additional support wasprovided by Intel. Views and conclusions contained in thisdocument are those of the authors and should not be inter-preted as representing the official policies, either expressedor implied, of DARPA, NSF, Intel or the U.S. government.Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for pro£t or commercial advantage and that copiesbear this notice and the full citation on the £rst page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior speci£cpermission and/or a fee.SIGCOMM’01, August 27-31, 2001, San Diego, California, USA..Copyright 2001 ACM 1-58113-411-8/01/0008 ...$5.00.1. INTRODUCTIONOver the last decade, researchers have studied how groupcommunication applications like audio and video conferenc-ing, multi-party games, content distribution, and broadcast-ing can be supported using IP Multicast[4]. However, overten years after its initial proposal, IP Multicast is yet tobe widely deployed due to fundamental concerns related toscalability, and support for higher layer functionality likereliability and congestion control. Recently, there has beena reevaluation by the research community of whether IP isindeed the right layer to support multicast-routing relatedfunctionality. A growing number of researchers [2, 3, 6, 9]have advocated an alternate architecture, where all multi-cast related functionality, including group management andpacket replication, is implemented at end systems. We referto such an architecture as End System Multicast. In thisarchitecture, end systems participating in a multicast groupself-organize into an overlay structure using a completelydistributed protocol. Further, end systems attempt to op-timize the efficiency of the overlay by adapting to networkdynamics and considering application level performance.While End System Multicast can have several potentialadvantages, a key concern is the performance of such anapproach. While several recent works have demonstratedthat the performance penalty of using overlays can be ac-ceptably low, these studies have been conducted primarilyusing simulation experiments, static metrics and controlledenvironments [2, 3, 9]. However, Internet environments, thetarget of these overlay architectures, are dynamic, heteroge-neous and unpredictable. In this paper we focus on a keyquestion regarding the feasibility of an overlay architecture:can such an architecture satisfy the demanding end-to-endperformance requirements of real world applications in suchan environment?We study this question in the context of an importantclass of applications: audio and video conferencing. Inter-net based conferencing applications have received a greatamount of attention in the last decade, during which ex-cellent tools like vic[10], vat[8] and rat[7] were developed.Yet, these tools are not ubiquitously deployed today due tothe limited availability of IP Multicast. Conferencing ap-plications have stringent performance requirements, and areamong the most challenging to support. They not only re-quire a high sustained throughput between the source andreceivers, but also require low latencies.In order to meet these performance requirements, we showthat it is necessary for self-organizing protocols to adapt toboth latency and bandwidth metrics. We present techniquesby which such protocols can adapt to dynamic metrics likeavailable bandwidth and latency, and yet remain resilient to55Figure 1: Example of End System Multicastnetwork noise and inaccuracies inherent in the measurementof these quantities. We demonstrate our ideas by incorpo-rating them into Narada, a self-organizing protocol that wepresented in [3]. While we have chosen to use Narada, webelieve that the techniques we present can be incorporatedinto other self-organizing protocols. In addition, althoughour techniques take advantage of some of the unique char-acteristics of conferencing applications, we believe that theycould benefit other classes of group communication applica-tions as well.We evaluate our techniques by testing the redesigned Nar-ada protocol on a wide-area test-bed. Our test-bed com-prises twenty machines that are distributed around NorthAmerica, Asia and Europe. Our results demonstrate thatour techniques can provide good performance, both