CONTENTS Message from MMTC Chair

Application Layer Multicast (ALM) is apromising technology for constructing contentdelivery systems across the Internet with lowcost and scalability. Unlike the conventional,well-known IP Multicast paradigm, ALM usesunicast to virtually realize multicastcommunications. In this article, we present asurvey of state-of-the-art techniques fordesigning efficient ALM trees and the challengeswe are facing. BackgroundAs opposed to IP Multicast (Fig. 1), in whichrouters play the role of data copying andforwarding at the network layer, ALM composesmulticast trees at the application layer (Fig. 2).Basically, there are two types of nodes in ALM,namely, a parent node and a leaf node. Parentnodes are supposed to take the role of routers inIP Multicast. The greatest benefit of ALM is thatusers can easily design their own contentdelivery systems without the constraints of otherlayers. ALM leans on the fact that the bandwidthof networks and the processing speed of endnodes have increased tremendously in recentyears, and this trend is sustainable in theforeseeable future.In ALM, the content server is treated as theroot and user nodes are considered as eitherparent nodes or leaf nodes depending on theirlocations. By making use of this tree-likestructure, ALM allows users to share thecontents in a multicast manner. ALM exhibitsmany attractive features such as scalability andsimplicity of using only unicast, yet severalissues need to be addressed. The first one is theaccumulation of delay at the lower layers inALM trees. The second is the degradation ofQoS caused by frequent joining and seceding ofnodes, especially the secession of nodes in theupper layers. The third, which impacts greatlywhether quality video streaming can be achieved,is the diversity of networks between the contentserver and receivers (Fig. 3). These three issueshave to be addressed in designing a contentdelivery system based on ALM.Existing ProposalsAiming at achieving high throughput and Qualityof Service (QoS), several ideas have beenproposed so far. From the view point of a treestructure, they can be classified into twocategories, namely, the so called single-treemulticast and multi-tree multicast, respectively. (1) Single-tree multicast (Fig.4) In the single-tree multicast, the server sendsstreaming traffic via a singular tree. Yoid [1],SpreadIt [2], ALMI [3], HBM [4], NICE [5],ZIGZAG [6], and Scribe [7] are examples thatfall into this category. Yoid [1] and SpreadIt [2]use the Shortest Path Tree to minimize the delayfrom the server to end nodes, whereas ALMI [3]and HBM [4] use the Minimum Spanning Treeto reduce the overall delay, both without theconsideration of bandwidth constraints. As awhole, the objective of these four methods is tominimize the content streaming delivery delay.NICE [5] and ZIGZAG [6] construct the tree byclustering. The purposes of using clustering areto curtail signaling overhead and to speed up treemanagement. Scribe [7] first uses Pastry [8] toconstruct the search paths and then makes thedelivery tree by tracing back the search paths.Scribe exhibits the merit of limiting the overheadof control messages, but without consideration ofthe available bandwidth of participating nodes. (2) Multiple-tree multicast The inability of single-tree multicast to facilitatenode secession is its main drawback. Multiple-tree multicast has thus been proposed toovercome this problem. Multiple-tree multicastadopts Multiple Description Coding (MDC)[9],[10] to first divide the original stream intomultiple descriptions, and then construct thesingle-tree multicast for each description,separately. The size of each description is, ingeneral, much smaller than that of the originalstream. Thus, playback is executable uponreceiving any description of the original stream.The more the number of descriptions received,the better the stream quality. MDC is now beingstudied widely for practical deployments such asincorporating MDC in the video coding standard IEEE COMSOC MMTC E-Letter http://www.comsoc.org/~mmc/15/34 Vol.4, No.8, September 2009H.264/AVC [11]-[13]. Multiple-tree multicastcan also prevent nodes from suffering theinterruption caused by the secession of nodes.CoopNet [14],[15], SplitStream [16], and THAG[17] are some representative methods that fallinto the multiple-tree multicast category. InCoopNet, all trees are managed at the server.Therefore, overloading the server might be aconcern. On the contrary, SplitStream [16]makes use of Scribe [7] to construct the trees.Trees in SplitStream are constructed in adistributed manner so that nodes can freely join atree at any position. These two methods haveaddressed the available bandwidth issue, but theydo not guarantee the node-disjoint property.Consequently, a node secession from the treemay result in simultaneous breakdown inmultiple trees. Incidentally, as shown in Fig. 5,the node-disjoint structure allows a parent nodein a certain tree to be a leaf node in the othertrees. In THAG, the node-disjoint structure isimplemented by the hierarchical ArrangementGraph (AG) [18], [19]. As a result, THAG ismore robust in terms of node secession ascompared to SplitStream or CoopNet. However,the main concern of THAG is that it does nottake network heterogeneity into consideration. Fig. 1. IP multicast