IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO.

1, FEBRUARY 2011

55

Live Streaming With Receiver-Based Peer-Division Multiplexing

Hyunseok Chang, Sugih Jamin, and Wenjie Wang
AbstractA number of commercial peer-to-peer (P2P) systems for live streaming have been introduced in recent years. The behavior of these popular systems has been extensively studied in several measurement papers. Due to the proprietary nature of these commercial systems, however, these studies have to rely on a black-box approach, where packet traces are collected from a single or a limited number of measurement points, to infer various properties of trafc on the control and data planes. Although such studies are useful to compare different systems from the end-users perspective, it is difcult to intuitively understand the observed properties without fully reverse-engineering the underlying systems. In this paper, we describe the network architecture of Zattoo, one of the largest production live streaming providers in Europe at the time of writing, and present a large-scale measurement study of Zattoo using data collected by the provider. To highlight, we found that even when the Zattoo system was heavily loaded with as high as 20 000 concurrent users on a single overlay, the median channel join delay remained less than 25 s, and that, for a majority of users, the streamed signal lags over-the-air broadcast signal by no more than 3 s. Index TermsLive streaming, network architecture, peer-topeer (P2P) system.

I. INTRODUCTION HERE is an emerging market for IPTV. Numerous commercial systems now offer services over the Internet that are similar to traditional over-the-air, cable, or satellite TV. Live television, time-shifted programming, and content-on-demand are all presently available over the Internet. Increased broadband speed, growth of broadband subscription base, and improved video compression technologies have contributed to the emergence of these IPTV services. We draw a distinction between three uses of peer-to-peer (P2P) networks: delay-tolerant le download of archival material, delay-sensitive progressive download (or streaming) of archival material, and real-time live streaming. In the rst case, the completion of download is elastic, depending on available bandwidth in the P2P network. The application buffer receives
Manuscript received December 25, 2009; accepted June 06, 2010; approved by IEEE/ACM TRANSACTIONS ON NETWORKING Editor P. Rodriguez. Date of publication August 03, 2010; date of current version February 18, 2011. H. Chang was with Zattoo, Inc., Ann Arbor, MI 48105 USA. He is now with Alcatel-Lucent Bell Labs, Holmdel, NJ 07733 USA (e-mail: hyunseok. chang@alcatel-lucent.com). S. Jamin is with the Electrical Engineering and Computer Science Department, University of Michigan, Ann Arbor, MI 48109 USA (e-mail: jamin@eecs. umich.edu). W. Wang was with Zattoo, Inc., Ann Arbor, MI 48105 USA. He is now with CRL, IBM Ressearch, Beijing 100193, China (e-mail: wenjwang@cn.ibm. com). Color versions of one or more of the gures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identier 10.1109/TNET.2010.2056382

data as it trickles in and informs the user upon the completion of download. The user can then start playing back the le for viewing in the case of a video le. Bittorrent and variants are examples of delay-tolerant le download systems. In the second case, video playback starts as soon as the application assesses it has sufcient data buffered that, given the estimated download rate and the playback rate, it will not deplete the buffer before the end of le. If this assessment is wrong, the application would have to either pause playback and rebuffer or slow down playback. While users would like playback to start as soon as possible, the application has some degree of freedom in trading off playback start time against estimated network capacity. Most video-on-demand systems are examples of delay-sensitive progressive-download application. The third case, real-time live streaming, has the most stringent delay requirement. While progressive download may tolerate initial buffering of tens of seconds or even minutes, live streaming generally cannot tolerate more than a few seconds of buffering. Taking into account the delay introduced by signal ingest and encoding, and network transmission and propagation, the live streaming system can introduce only a few seconds of buffering time end-to-end and still be considered live [1]. The Zattoo peer-to-peer live streaming system was a free-to-use network serving over 3 million registered users in eight European countries at the time of study, with a maximum of over 60 000 concurrent users on a single channel. The system delivers live streams using a receiver-based, peer-division multiplexing scheme as described in Section II. To ensure real-time performance when peer uplink capacity is below requirement, Zattoo subsidizes the networks bandwidth requirement, as described in Section III. After delving into Zattoos architecture in detail, we study in Sections IV and V large-scale measurements collected during the live broadcast of the UEFA European Football Championship, one of the most popular one-time events in Europe, in June 2008 [2]. During the course of that month, Zattoo served more than 35 million sessions to more than 1 million distinct users. Drawing from these measurements, we report on the operational scalability of Zattoos live streaming system along several key issues. 1) How does the system scale in terms of overlay size and its effectiveness in utilizing peers uplink bandwidth? 2) How responsive is the system during channel switching, for example, when compared to the 3-s channel switch time of satellite TV? 3) How effective is the packet retransmission scheme in allowing a peer to recover from transient congestion? 4) How effective is the receiver-based peer-division multiplexing scheme in delivering synchronized substreams?

1063-6692/$26.00 2010 IEEE

56

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

Fig. 1. Zattoo delivery network architecture.

5) How effective is the global bandwidth subsidy system in provisioning for ash crowd scenarios? 6) Would a peer farther away from the stream source experience adversely long lag compared to a peer closer to the stream source? 7) How effective is error-correcting code in isolating packet losses on the overlay? We also discuss in Section VI several challenges in increasing the bandwidth contribution of Zattoo peers. Finally, we describe related works in Section VII and conclude in Section VIII. II. SYSTEM ARCHITECTURE The Zattoo system rebroadcasts live TV, captured from satellites, onto the Internet. The system carries each TV channel on a separate peer-to-peer delivery network and is not limited in the number of TV channels it can carry. Although a peer can freely switch from one TV channel to another, thereby departing and joining different peer-to-peer networks, it can only join one peer-to-peer network at any one time. We henceforth limit our description of the Zattoo delivery network as it pertains to carrying one TV channel. Fig. 1 shows a typical setup of a single TV channel carried on the Zattoo network. TV signal captured from satellite is encoded into H.264/AAC streams, encrypted, and sent onto the Zattoo network. The encoding server may be physically separated from the server delivering the encoded content onto the Zattoo network. For ease of exposition, we will consider the two as logically colocated on an Encoding Server. Users are required to register themselves at the Zattoo Web site to download a free copy of the Zattoo player application. To receive the signal of a channel, the user rst authenticates itself to the Zattoo Authentication Server. Upon authentication, the user is granted a ticket with limited lifetime. The user then presents this ticket, along with the identity of the TV channel of interest, to the Zattoo Rendezvous Server. If the ticket species that the user is authorized to receive signal of the said TV channel, the Rendezvous Server returns to the user a list of peers currently joined to the P2P network carrying the channel, together with a signed channel ticket. If the user is the rst peer to join a channel, the list of peers it receives contain only the Encoding Server. The user joins the channel by contacting the peers returned by the Rendezvous Server, presenting its channel

ticket, and obtaining the live stream of the channel from them (see Section II-A for details). Each live stream is sent out by the Encoding Server as logical substreams. The signal received from satellite is encoded into a variable-bit rate stream. During periods of source quiescence, no data is generated. During source busy periods, generated data is packetized into a packet stream, with each packet limited to a maximum size. The Encoding Server multiplexes this packet stream onto the Zattoo network as logical substreams. Thus, the rst packet generated is considered part of the rst substream, the second packet that of the second sub-th stream, the th packet that of the th substream. The packet cycles back to the rst substream, etc., such that the th substream carries the -th packets, where , , and a user-dened constant. We call a set of packets with the same index multiplier a segment. Thus serves as the segment index, while serves as the packet index within a segment. Each segment is of size packets. Being the packet index, also serves as the substream index. The number is carried in each packet as its sequence number. Zattoo uses the ReedSolomon (RS) error correcting code (ECC) for forward error correction. The RS code is a systematic packets carry code: of the packets sent per segment, the live stream data, while the remainder carries the redundant data [3, Sec. 7.3]. Due to the variable-bit rate nature of the data stream, the time period covered by a segment is variable, and a packet may be of size less than the maximum packet size. A packet smaller than the maximum packet size is zero-padded to the maximum packet size for the purposes of computing the (shortened) RS code, but is transmitted in its original size. Once a peer has received packets per segment, it can reconstruct the remaining packets. We do not differentiate between streaming data and redundant data in our discussion in the remainder of this paper. When a new peer requests to join an existing peer, it species the substream(s) it would like to receive from the existing peer. These substreams do not have to be consecutive. Contingent upon availability of bandwidth at existing peers, the receiving peer decides how to multiplex a stream onto its set of neighboring peers, giving rise to our description of the Zattoo live streaming protocol as a receiver-based, peer-division multiplexing protocol. The details of peer-division multiplexing is described in Section II-A, while the details of how a peer manages substream forwarding and stream reconstruction is described in Section II-B. Receiver-based peer-division multiplexing has also been used by the latest version of CoolStreaming peer-topeer protocol, though it differs from Zattoo in its stream management (Section II-B) and adaptive behavior (Section II-C) [4]. A. Peer-Division Multiplexing To minimize per-packet processing time of a stream, the Zattoo protocol sets up a virtual circuit with multiple fanouts at each peer. When a peer joins a TV channel, it establishes a peer-division multiplexing (PDM) scheme among a set of neighboring peers by building a virtual circuit to each of the neighboring peers. Baring departure or performance degradation of a neighbor peer, the virtual circuits are maintained until the joining peer switches to another TV channel. With the

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

57

virtual circuits set up, each packet is forwarded without further per-packet handshaking between peers. We describe the PDM boot-strapping mechanism in this section, and the adaptive PDM mechanism to handle peer departure and performance degradation in Section II-C. The PDM establishment process consists of two phases: the search phase and the join phase. In the search phase, the new joining peer determines its set of potential neighbors. In the join phase, the joining peer requests peering relationships with a subset of its potential neighbors. Upon acceptance of a peering relationship request, the peers become neighbors, and a virtual circuit is formed between them. 1) Search Phase: To obtain a list of potential neighbors, a joining peer sends out a SEARCH message to a random subset of the existing peers returned by the Rendezvous Server. The SEARCH message contains the substream indices for which this joining peer is looking for peering relationships. The substream indices are usually represented as a bitmask of bits, where is the number of substreams dened for the TV channel. In the beginning, the joining peer will be looking for peering relationships for all substreams and have all the bits in the bitmask turned on. In response to a SEARCH message, an existing peer replies with the number of substreams it can forward. From the returning SEARCH replies, the joining peer constructs a set of potential neighbors that covers the full set of substreams comprising the live stream of the TV channel. The joining peer continues to wait for SEARCH replies until the set of potential neighbors contains at least a minimum number of peers, or until all SEARCH replies have been received. With each SEARCH reply, the existing peer also returns a random subset of its known peers. If a joining peer cannot form a set of potential neighbors that covers all of the substreams of the TV channel, it initiates another SEARCH round, sending SEARCH messages to peers newly learned from the previous round. The joining peer gives up if it cannot obtain the full stream after two SEARCH rounds. To help the joining peer synchronize the substreams it receives from multiple peers, each existing peer also indicates for each substream the latest sequence number it has received for that substream and the existence of any quality problem. The joining peer can then choose substreams with good quality that are closely synchronized. 2) Join Phase: Once the set of potential neighbors is established, the joining peer sends JOIN requests to each potential neighbor. The JOIN request lists the substreams for which the joining peer would like to construct virtual circuit with the potential neighbor. If a joining peer has potential neighbors, each willing to forward it the full stream of a TV channel, it would of the stream to typically choose to have each forward only spread out the load among the peers and to speed up error recovery, as described in Section II-C. In selecting which of the potential neighbors to peer with, the joining peer gives highest preference to topologically close-by peers, even if these peers have less capacity or carry lower quality substreams. The topological location of a peer is dened to be its subnet number, autonomous system (AS) number, and country code, in that order of precedence. A joining peer obtains its own topological location from the Zattoo Authentication Server as part of its authentication process. The list of peers returned by both the Ren-

Fig. 2. Zattoo peer with IOB.

dezvous Server and potential neighbors all come attached with topological locations. A topology-aware overlay not only allows us to be ISP-friendly, by minimizing interdomain trafc and thus saving on transit bandwidth cost, but also helps reduce the number of physical links and metro hops traversed in the overlay network, potentially resulting in enhanced user-perceived stream quality. B. Stream Management We represent a peer as a packet buffer, called the IOB, fed by substreams incoming from the PDM constructed as described in Section II-A.1 The IOB drains to : 1) a local media player if one is running; 2) a local le if recording is supported; and 3) potentially other peers. Fig. 2 depicts a Zattoo player application with virtual circuits established to four peers. As packets from each substream arrive at the peer, they are stored in the IOB for reassembly to reconstruct the full stream. Portions of the stream that have been reconstructed are then played back to the user. In addition to providing a reassembly area, the IOB also allows a peer to absorb some variabilities in available network bandwidth and network delay. The IOB is referenced by an input pointer, a repair pointer, and one or more output pointers. The input pointer points to the slot in the IOB where the next incoming packet with sequence number higher than the highest sequence number received so far will be stored. The repair pointer always points one slot beyond the last packet received in order and is used to regulate packet retransmission and adaptive PDM as described later. We assign an output pointer to each forwarding destination. The output pointer of a destination indicates the destinations current forwarding horizon on the IOB. In accordance to the three types
1In the case of the Encoding Server, which we also consider a peer on the Zattoo network, the buffer is fed by the encoding process.

58

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

Fig. 3. Packet map associated with a peer pointer.

of possible forwarding destinations listed, we have three types of output pointers: player pointer, le pointer, and peer pointer. One would typically have at most one player pointer and one le pointer, but potentially multiple concurrent peer pointers, referencing an IOB. The Zattoo player application does not currently support recording. Since we maintain the IOB as a circular buffer, if the incoming packet rate is higher than the forwarding rate of a particular destination, the input pointer will overrun the output pointer of that destination. We could move the output pointer to match the input pointer so that we consistently forward the oldest packet in the IOB to the destination. Doing so, however, requires checking the input pointer against all output pointers on every packet arrival. Instead, we have implemented the IOB as a double buffer. With the double buffer, the positions of the output pointers are checked against that of the input pointer only when the input pointer moves from one sub-buffer to the other. When the input pointer moves from sub-buffer to sub-buffer , all the output pointers still pointing to sub-buffer are moved to the start of sub-buffer and sub-buffer is ushed, ready to accept new packets. When a sub-buffer is ushed while there are still output pointers referencing it, packets that have not been forwarded to the destinations associate with those pointers are lost to them, resulting in quality degradation. To minimize packet lost due to sub-buffer ushing, we would like to use large sub-buffers. However, the real-time delay requirement of live streaming limits the usefulness of late arriving packets and effectively puts a cap on the maximum size of the sub-buffers. Different peers may request for different numbers of, possibly nonconsecutive, substreams. To accommodate the different forwarding rates and regimes required by the destinations, we associate a packet map and forwarding discipline with each output pointer. Fig. 3 shows the packet map associated with an output peer pointer where the peer has requested substreams 1, 4, 9, and 14. Every time a peer pointer is repositioned to the beginning of a sub-buffer of the IOB, all the packet slots of the requested substreams are marked NEEDed and all the slots of the substreams not requested by the peer are marked SKIP. When a NEEDed packet arrives and is stored in the IOB, its state in the packet map is changed to READY. As the peer pointer moves along its associated packet map, READY packets are forwarded to the peer and their states changed to SENT. A slot in Fig. 3, marked NEEDed but not READY, such as slot indicates that the packet is lost or will arrive out of order and is bypassed. When an out-of-order packet arrives, its slot is changed to READY, and the peer pointer is reset to point to this slot. Once the out-of-order packet has been sent to the peer, the peer pointer will move forward, bypassing all SKIP, NEED, and

Fig. 4. IOB, input/output pointers, and packet maps.

SENT slots until it reaches the next READY slot, where it can resume sending. The player pointer behaves the same as a peer pointer except that all packets in its packet map will always start out marked NEEDed. Fig. 4 shows an IOB consisting of a double buffer, with an input pointer, a repair pointer, and an output le pointer, an output player pointer, and two output peer pointers referencing the IOB. Each output pointer has a packet map associated with it. For the scenario depicted in the gure, the player pointer tracks the input pointer and has skipped over some lost packets. Both peer pointers are lagging the input pointer, indicating that the forwarding rates to the peers are bandwidth-limited. The le pointer is pointing at the rst lost packet. Archiving a live stream to le does not impose real-time delay bound on packet arrivals. To achieve the best quality recording possible, a recording peer always waits for retransmission of lost packets that cannot be recovered by error correction. In addition to achieving lossless recording, we use retransmission to let a peer recover from transient network congestion. A peer sends out a retransmission request when the distance between the repair pointer and the input pointer has reached a threshold of packet slots, usually spanning multiple segments. A retransmission request consists of an -bit packet mask, with each bit representing a packet, and the sequence number of the packet corresponding to the rst bit. Marked bits in the packet mask indicate that the corresponding packets need to be retransmitted. When a packet loss is detected, it could be caused by congestion on the virtual circuits forming the current PDM or congestion on the path beyond the neighboring peers. In either case, current neighbor peers will not be good sources of retransmitted packets. Hence, we send our retransmission requests to random peers that are not neighbor peers. A peer receiving a retransmission request will honor the request only if the requested packets are still in its IOB and it has sufcient leftover capacity, after serving its current peers, to transmit all the requested packets. Once a retransmission request is accepted, the peer will retransmit all the requested packets to completion. C. Adaptive PDM While we rely on packet retransmission to recover from transient congestions, we have two channel capacity adjustment

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

59

mechanisms to handle longer-term bandwidth uctuations. The rst mechanism allows a forwarding peer to adapt the number of substreams it will forward given its current available bandwidth, while the second allows the receiving peer to switch provider at the substream level. Peers on the Zattoo network can redistribute a highly variable number of substreams, reecting the high variability in uplink bandwidth of different access network technologies. For a full-stream consisting of 16 constant-bit rate substreams, our prior study showed that based on realistic peer characteristics measured from the Zattoo network, half of the peers can support less than half of a stream, 82% of peers can support less than a full-stream, and the remainder can support up to 10 full streams (peers that can redistribute more than a full stream is conventionally known as supernodes in the literature) [5]. With variable-bit rate streams, the bandwidth carried by each substream is also variable. To increase peer bandwidth usage, without undue degradation of service, we instituted measurement-based admission control at each peer. In addition to controlling resource commitment, another goal of the measurement-based admission control module is to continually estimate the amount of available uplink bandwidth at a peer. The amount of available uplink bandwidth at a peer is initially estimated by the peer sending a pair of probe packets to Zattoos Bandwidth Estimation Server. Once a peer starts forwarding substreams to other peers, it will receive from those peers quality-of-service feedbacks that inform its update of available uplink bandwidth estimate. A peer sends quality-of-service feedback only if the quality of a substream drops below a certain threshold.2 Upon receiving quality feedback from multiple peers, a peer rst determines if the identied substreams are arriving in low quality. If so, the low quality of service may not be caused by limit on its own available uplink bandwidthin which case, it ignores the low quality feedbacks. Otherwise, the peer decrements its estimate of available uplink bandwidth. If the new estimate is below the bandwidth needed to support existing number of virtual circuits, the peer closes a virtual circuit. To reduce the instability introduced into the network, a peer closes rst the virtual circuit carrying the smallest number of substreams. A peer attempts to increase its available uplink bandwidth estimate periodically if it has fully utilized its current estimate of available uplink bandwidth without triggering any bad quality feedback from neighboring peers. A peer doubles the estimated available uplink bandwidth if current estimate is below a threshold, switching to linear increase above the threshold, similar to how TCP maintains its congestion window size. A peer also increases its estimate of available uplink bandwidth if a neighbor peer departs the network without any bad quality feedback. When the repair pointer lags behind the input pointer by packet slots, in addition to initiating a retransmission request, a
2Depending on a peers NAT and/or rewall conguration, Zattoo uses either UDP or TCP as the underlying transport protocol. The quality of a substream is measured differently for UDP and TCP. A packet is considered lost under UDP if it does not arrive within a xed threshold. The quality measure for UDP is computed as a function of both the packet lost rate and the burst error rate (number of contiguous packet losses). The quality measure for TCP is dened to be how far behind a peer is, relative to other peers, in serving its substreams.

peer also computes a loss rate over the packets. If the loss rate is above a threshold, the peer considers the neighbor slow and attempts to recongure its PDM. In reconguring its PDM, a peer attempts to shift half of the substreams currently forwarded by the slow neighbor to other existing neighbors. At the same time, it searches for new peer(s) to forward these substreams. If new peer(s) are found, the load will be shifted from existing neighbors to the new peer(s). If substreams from the slow neighbor continue to suffer after the reconguration of the PDM, the peer will drop the neighbor completely and initiate another reconguration of the PDM. When a peer loses a neighbor due to reduced available uplink bandwidth at the neighbor or due to neighbor departure, it also initiates a PDM reconguration. A peer may also initiate a PDM reconguration to switch to a topologically closer peer. Similar to the PDM establishment process, PDM reconguration is accomplished by peers exchanging substream bitmasks in a request/response handshake, with each bit of the bitmask representing a substream. During and after a PDM reconguration, slow neighbor detection is disabled for a short period of time to allow for the system to stabilize. III. GLOBAL BANDWIDTH SUBSIDY SYSTEM Each peer on the Zattoo network is assumed to serve a user through a media player, which means that each peer must receive, and can potentially forward, all substreams of the TV channel the user is watching. The limited redistribution capacity of peers on the Zattoo network means that a typical client can contribute only a fraction of the substreams that make up a channel. This shortage of bandwidth leads to a global bandwidth decit in the peer-to-peer network. Whereas bittorrentlike delay-tolerant le downloads or the delay-sensitive progressive download of video-on-demand applications can mitigate such global bandwidth shortage by increasing download time, a live streaming system such as Zattoos must subsidize the bandwidth shortfall to provide real-time delivery guarantee. Zattoos Global Bandwidth Subsidy System (or simply, the Subsidy System), consists of a global bandwidth monitoring subsystem, a global bandwidth forecasting and provisioning subsystem, and a pool of Repeater nodes. The monitoring subsystem continuously monitors the global bandwidth requirement of a channel. The forecasting and provisioning subsytem projects global bandwidth requirement based on measured history and allocates Repeater nodes to the channel as needed. The monitoring and provisioning of global bandwidth is complicated by two highly varying parameters over time, client population size and peak streaming rate, and one varying parameter over space, available uplink bandwidth, which is network service provider-dependent. Forecasting of bandwidth requirement is a vast subject in itself. Zattoo adopted a very simple mechanism, described in Section III-B, which has performed adequately in provisioning the network for both daily demand uctuations and ash crowds scenarios (see Section IV-C). When a bandwidth shortage is projected for a channel, the Subsidy System assigns one or more Repeater nodes to the channel. Repeater nodes function as bandwidth multiplier, to amplify the amount of available bandwidth in the network. Each Repeater node serves at most one channel at a time; it

60

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

joins and leaves a given channel at the behest of the Subsidy System. Repeater nodes receive and serve all substreams of the channel they join, run the same PDM protocol, and are treated by actual peers like any other peers on the network. However, as bandwidth ampliers, they are usually provisioned to contribute more uplink bandwidth than the download bandwidth they consume. The use of Repeater nodes makes the Zattoo network a hybrid P2P and content distribution network. We next describe the bandwidth monitoring subsystem of the Subsidy System, followed by design of the simple bandwidth projection and Repeater node assignment subsystem. A. Global Bandwidth Measurement The capacity metric of a channel is the tuple , where is the aggregate download rates required by all users on the channel, and is the aggregate upload capacity of those users. , and the difference between the two is the Usually global bandwidth decit of the channel. Since channel viewership changes over time as users join or leave the channel, we need a scalable means to measure and update the capacity metric. We rely on aggregating the capacity metric up the peerdivision multiplexing tree. Each peer in the overlay periodically aggregates the capacity metric reported by all its downstream to the agreceiver peers, adds its own capacity measure gregate, and forwards the resulting capacity metric upstream to its forwarding peers. By the time the capacity metric percolates up to the Encoding Server, it contains the total download and upload rate aggregates of the whole streaming overlay. The Ento the coding Server then simply forwards the obtained Subsidy Server. B. Global Bandwidth Projection and Provisioning For each channel, the Subsidy Server keeps a history of the capacity metric reports received from the channels Encoding Server. The channel utilization ratio ( ) is the ratio over . Based on recent movements of the ratio , we classify the capacity trend of each channel into the following four categories. for Stable: The ratio has remained within the past reporting periods. Exploding: The ratio increased by at least between (e.g., ) reporting periods. Increasing: The ratio has steadily increased by ( ) over the past reporting periods. Falling: The ratio has decreased by over the past reporting periods. Orthogonal to the capacity-trend-based classication, each channel is further categorized in terms of its capacity utilization ratio as follows. Underutilized: The ratio is below the low threshold , . e.g., Near Capacity: The ratio is almost 1.0, e.g., . If the capacity trend of a channel has been Exploding, one or more Repeater nodes will be assigned to it immediately. If the channels capacity trend has been Increasing, it will be assigned a Repeater node with a smaller capacity. The goal of the Subsidy System is to keep a channel below Near Capacity. If a

TABLE I SESSION DATABASE (JUNE 130, 2008)

channels capacity trend is Stable and the channel is Underutilized, the Subsidy System attempts to free Repeater nodes (if any) assigned to the channel. If a channels capacity utilization is Falling, the Subsidy System waits for the utilization to stabilize before reassigning any Repeater nodes. Each Repeater node periodically sends a keep-alive message to the Subsidy System. The keep-alive message tells the Subsidy System which channel the Repeater node is serving, plus its CPU and capacity utilization ratio. This allows the Subsidy System to monitor the health of Repeater nodes and to increase the stability of the overlay during Repeater reassignment. When reassigning Repeater nodes from a channel, the Subsidy System will start from the Repeater node with the lowest utilization ratio. It will notify the selected Repeater node to stop accepting new peers and then to leave the channel after a specied time. In addition to Repeater nodes, the Subsidy System may recognize extra capacity from idle peers whose owners are not actively watching a channel. However, our previous study shows that a large number of idle peers are required to make any discernible impact on the global bandwidth decit of a channel [5]. Our current Subsidy System therefore does not solicit bandwidth contribution from idle peers. IV. SERVER-SIDE MEASUREMENTS In the Zattoo system, two separate centralized collector servers collect usage statistics and error reports, which we call the stats server and the user-feedback server respectively. The stats server periodically collects aggregated player statistics from individual peers, from which full session logs are constructed and entered into a session database. The session database gives a complete picture of all past and present sessions served by the Zattoo system. A given database entry contains statistics about a particular session, which includes join time, leave time, uplink bytes, download bytes, and channel name associated with the session. We study the sessions generated on three major TV channels from three different countries (Germany, Spain, and Switzerland), from June 130, 2008. Throughout the paper, we label those channels from Germany, Spain, and Switzerland as ARD, Cuatro, and SF2, respectively. Euro 2008 games were held during this period, and those three channels broadcast a majority of the games, including the nal match. See Table I for information about the collected session data sets. The user-feedback server, on the other hand, collects users error logs submitted asynchronously by users. The user feedback data here is different from peers quality feedback used in PDM reconguration described in Section II-C. Zattoo player maintains an encrypted log le that contains, for debugging purposes, detailed behavior of client-side P2P engine, as well as history of all the streaming sessions initiated by a user since the player startup. When users encounter any error while using the

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

61

Fig. 5. Overlay size and sharing ratio. (a) ARD. (b) Cuatro. (c) SF2.

TABLE II FEEDBACK LOGS (JUNE 2029, 2008)

TABLE III AVERAGE SHARING RATIO

player, such as login error, join failure, bad quality streaming, etc., they can choose to report the error by clicking a Submit Feedback button on the player, which causes the Zattoo player to send the generated log le to the user-feedback server. Since a given feedback log not only reports on a particular error, but also describes normal sessions generated prior to the occurrence of the error, we can study the users viewing experience (e.g., channel join delay) from the feedback logs. Table II describes the feedback logs collected from June 2029. A given feedback log can contain multiple sessions (for the same or different channels), depending on users viewing behavior. The second column in the table represents the number of feedback logs that contain at least one session generated on the channel listed in the corresponding entry in the rst column. The numbers in the third column indicate how many distinct sessions generated on said channel are present in the feedback logs. A. Overlay Size and Sharing Ratio We rst study how many concurrent users are supported by the Zattoo system and how much bandwidth is contributed by them. For this purpose, we use the session database presented in Table I. By using the join/leave timestamps of the collected sessions, we calculate the number of concurrent users on a given channel at time . Then, we calculate the average sharing ratio of the given channel at the same time. The average sharing ratio is dened as total users uplink rate divided by their download rate on the channel. A sharing ratio of one means users contribute to other peers in the network as much trafc as they download at the time. We calculate the average sharing ratio from the total download/uplink bytes of the collected sessions. We rst obtain all the sessions that are active across time . We call a set of such sessions . Then, assuming uplink/download bytes of each session are spread uniformly throughout the entire session duration, we approximate the average sharing ratio at time as

Fig. 5 shows the overlay size (i.e., number of concurrent users) and average sharing ratio superimposed across the month of June 2008. According to the gure, the overlay size grew to more than 20 000 (e.g., 20 059 on ARD on June 18 and 22 152 on Cuatro on June 9). As opposed to the overlay size, the average sharing ratio tends to stay atter throughout the month. Occasional spikes in the sharing ratio all occurred during 26 a.m. (GMT) when the channel usage is very low and may therefore be considered statistically insignicant. By segmenting a 24-h day into two time periods, e.g., offpeak hours (12 a.m.-6 p.m.) and peak hours (6 p.m.-12 a.m.), Table III shows the average sharing ratio in the two time periods separately. Zattoos usage during peak hours typically accounts for about 50%70% of the total usage of the day. According to the table, the average sharing ratio during peak hours is slightly lower than, but not very much different from, during off-peak hours. Cuatro channel in Spain exhibits relatively lower sharing ratio than the two other channels. One bandwidth test site [6] reports that average uplink bandwidth in Spain is about 205 kb/s, which is much lower than in Germany (582 kb/s) and Switzerland (787 kb/s). The lower sharing ratio on the Spanish channel may reect regional difference in residential access network provisioning. The balance of the required bandwidth is provided by Zattoos Encoding Server and Repeater nodes. B. Channel Switching Delay When a user clicks on a new channel button, it takes some time (a.k.a. channel switching delay) for the user to be able to start watching streamed video on the Zattoo player. The channel switching delay has two components. First, the Zattoo player needs to contact other available peers and retrieve all required substreams from them. We call the delay incurred during this stage join delay. Once all necessary substreams have been negotiated successfully, the player then needs to wait and buffer the minimum amount of streams (e.g., 3 s) before starting to show the video to the user. We call the resulting

62

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

Fig. 6. CDF of user arrival time. (a) ARD. (b) Cuatro. (c) SF2.

Fig. 7. CDF of channel join delay. (a) ARD. (b) Cuatro. (c) SF2.

wait time buffering delay. The total channel switching delay experienced by users is thus the sum of join delay and buffering delay. PPLive reports channel switching delay around 20 to 30 s, but can be as high as 2 min, of which join delay typically accounts for 1015 s [7]. We measure the join delay experienced by Zattoo users from the feedback logs described in Table II. Debugging information contained in the feedback logs tells us when the user clicked on a particular channel, and when the player has successfully joined the P2P overlay and started to buffer content. One concern in relying on user-submitted feedback logs to infer join delay is the potential sampling bias associated with them. Users typically submit feedback logs when they encounter some kind of errors, and that brings up the question of whether the captured sessions are representative samples to study. We attempt to address this concern by comparing the data from feedback logs against those from the session database. The latter captures the complete picture of the users channel watching behavior and can therefore serve as a reference. In our analysis, we compare the user arrival time distribution obtained from the two data sets. For fair comparison, we used a subset of the session database that was generated during the same period when the feedback logs were collected (i.e., from June 2029). Fig. 6 plots the CDF distribution of user arrivals per hour obtained from feedback logs and session database separately. The steep slope of the distributions during hours 1820 (68 p.m.) indicates the high frequency of user arrivals during those hours. On ARD and Cuatro, the user arrival distributions inferred from feedback logs are almost identical to those from theh session database. On the other hand, on SF2, the distribution obtained

TABLE IV MEDIAN CHANNEL JOIN DELAY

from feedback logs tends to grow slowly during early hours, which indicates that feedback submission rate during off-peak hours on SF2 was relatively lower than normal. Later during peak hours, however, the feedback submission rate picks up as expected, closely matching the actual user arrival rate. Based on this observation, we argue that feedback logs can serve as representative samples of daily user activities. Fig. 7 shows the CDF distributions of channel join delay for ARD, Cuatro, and SF2 channels. We show the distributions for off-peak hours (12 a.m.-6 p.m.) and peak hours (6 p.m.-12 a.m.) separately. Median channel join delay is also presented in a similar fashion in Table IV. According to the CDF distributions, 80% of users experience less than 48 s of join delay, and 50% of users even less than 2 s of join delay. Also, Table IV shows that even a 10-fold increase on the number of concurrent users during peak hours does not unduly lengthen the channel join delay (up to 22% increase in median join delay). C. Repeater Node Assignment As illustrated in Fig. 5, the live coverage of Euro Cup games brought huge ash crowds to the Zattoo system. With typical users contributing only about 25%30% of the average

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

63

Fig. 8. Overlay size and channel provisioning. (a) ARD. (b) Cuatro. (c) SF2.

streaming bitrate, Zattoo must subsidize the balance. As described in Section III, the Zattoos Subsidy System assigns Repeater nodes to channels that require more bandwidth than its own globally available aggregate upload bandwidth. Fig. 8 shows the Subsidy System assigning more Repeater nodes to a channel as ash crowds arrived during each Euro Cup game, and then gradually reclaiming them as the ash crowd departed. For each of the channels reported, we choose a particular date with the biggest ash crowd on the channel. The dip in overlay sizes on ARD and Cuatro channels occurred during the half-time break of the games. The Subsidy Server was less aggressive in assigning Repeater nodes to the ARD channel, as compared to the other two channels, because the Repeater nodes in the vicinity of the ARD server have higher capacity than those near the other two. V. CLIENT-SIDE MEASUREMENTS To further study the P2P overlay beyond details obtainable from aggregated session-level statistics, we run several modied Zattoo clients that periodically retrieve the internal states of other participating peers in the network by exchanging SEARCH/JOIN messages with them. After a given probe session is over, the monitoring client archives a log le where we can analyze control/data trafc exchanged and detailed protocol behavior. We run the experiment during Zattoos live coverage of Euro 2008 (June 729). The monitoring clients tuned to game channels from one of Zattoos data centers located in Switzerland while the games were broadcast live. The data sets presented in this paper were collected during the coverage of the championship nal on two separate channels: ARD in Germany and Cuatro in Spain. Soccer teams from Germany and Spain participated in the championship nal. As described in Section II-A, Zattoos peer discovery is guided by the peers topology information. To minimize potential sampling bias caused by our use of a single vantage point for monitoring, we assigned empty AS number and country code to our monitoring clients, so that their probing is not geared toward those peers located in the same AS and country. A. Substream Synchrony To ensure good viewing quality, the peer should not only obtain all necessary substreams (discounting redundant substreams), but also have those substreams delivered temporally

synchronized with each other for proper online decoding. Receiving out-of-sync substreams typically results in a pixelated screen on the player. As described in Sections I and II-C, Zattoos protocol favors substreams that are relatively in sync when constructing the PDM, and continually monitors the substreams progression over time, replacing those substreams that have fallen behind and reconguring the PDM when necessary. In this section, we measure the effectiveness of Zattoos Adaptive PDM in selecting substreams that are largely in sync. To quantify such intersub stream synchrony, we measure the difference in the latest (i.e., maximum) packet sequence numbers belonging to different incoming substreams. When a remote peer responds to a SEARCH query message, it includes in its SEARCH reply the latest sequence numbers that it has received for all substreams. If some substreams happen to be lossy or stalled at that time, the peer marks such substreams in its SEARCH replies. Thus, we can inspect SEARCH replies from existing peers to study their intersubstream synchrony. In our experiment, we collected SEARCH replies from 4420 and 6530 distinct peers from ARD and Cuatro channels, respectively, during the 2-h coverage of the nal game. From the collected SEARCH replies, we check how many substreams (out of 16) are bad (e.g., lossy or missing) for each peer. Fig. 9(a) shows the CDF distribution of the number of bad substreams. According to the gure, about 99% (ARD) and 96% (Cuatro) of peers have three or less bad substreams. substreams (out Current Zattoo deployment dedicates ) for loss recovery purposes. That is, given a segof ment of 16 consecutive packets, if peer has received at least 13 packets, it can reconstruct the remaining three packets from the RS error-correcting code (see Section I). Thus, if the peer can receive any 13 substreams out of 16 reliably, it can decode the full stream properly. The result in Fig. 9(a) suggests that the number of bad substreams is low enough as to not cause quality issues in the Zattoo network. After discounting bad substreams, we then look at the synchrony of the remaining good substreams in each peer. Fig. 9(b) shows the CDF distribution of the substream synchrony in the two channels. Substream synchrony of a given peer is dened as the difference between maximum and minimum packet sequence numbers among all substreams, which is obtained from the peers SEARCH reply. For example, if some peer has substream synchrony measured at 100, it means that the peer has one substream that is ahead of another substream

64

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

Fig. 9. Sub-stream synchrony. (a) CDF for number of bad substreams. (b) CDF for substream synchrony.

by 100 packets. If all the packets are received in order, the . If substream synchrony of a peer measures at most we received multiple SEARCH replies from the same peer, we average the substream synchrony across all the replies. Given the 500-kb/s average channel data rate, 60 consecutive packets roughly correspond to 1 s worth of streaming. Thus, the gure shows that on Cuatro channel, 20% of peers have their substreams completely in sync, while more than 90% have their substreams lagging each other by at most 5 s. On ARD channel, 30% are in sync, and more than 90% are within 1.5 s. The buffer space of Zattoo player has been dimensioned sufciently to accommodate such low degree of out-of-sync substreams. B. Peer Synchrony While substream synchrony tells us stream quality different peers may experience, peer synchrony tells us how varied in time peers viewing points are. With small-scale P2P networks, all participating peers are likely to watch live streaming roughly synchronized in time. However, as the size of the P2P overlay grows, the viewing point of edge nodes may be delayed significantly compared to those closer to the Encoding Server. In the experiment, we dene the viewing point of a peer as the median of the latest sequence numbers across its substreams. Then, we choose one peer (e.g., a Repeater node directly connected to the Encoding Server) as a reference point and compare other peers viewing point to the reference viewing point. Fig. 10 shows the CDFs of relative peer synchrony. The relative peer synchrony of peer is obtained by the viewing point of subtracted by the reference viewing point. Peer synchrony at 60 means that a given peers viewing point is delayed by 60 packets (roughly 1 s for a 500-kb/s stream) compared to the reference viewing point. A positive viewing point means that a given peers stream gets ahead of the reference point, which could happen for peers that receive streams directly from the Encoding Server. The gure shows that about 1% of peers on ARD and 4% of peers on Cuatro experienced more than 3 s (i.e., 180 packets) delay in streaming compared to the reference viewing point. To better understand to what extent the peers position in the overlay affects peer synchrony, we plot in Fig. 11 the CDFs
Fig. 10. Peer synchrony.

of peer synchrony at different depths, i.e., distances from the Encoding Server. We look at how much delay is introduced for substreams traversing peers from the Encoding Server (i.e., depth ). For this purpose, we associate per-substream viewing point information from SEARCH reply with per-substream overlay depth information from JOIN reply, where SEARCH/JOIN replies were sent from the same peer close in time (e.g., less than 1 s apart). If the length of the peer-hop from the Encoding Server has an adverse effect on playback delay and viewing point, we expect peers farther away from the Encoding Server to be further offset from the reference viewing point, resulting in CDFs that are shifted to the upper left corner for peers at farther distances from the Encoding Server. The gure shows an absence of such a leftward shift in the CDFs. The median delay at depth 6 does not grow by more than 0.5 s compared to the median delay at depth 2. The gure shows that having more peers farther away from the Encoding Servers increases the number of peers that are 0.5 s behind the reference viewing point without increasing the offset in viewing point itself. On the Zattoo network, once the virtual circuits comprising a PDM have been set up, each packet is streamed with minimal delay through each peer. Hence, each peer-hop from the Encoding Server introduces delay only in the tens of

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

65

Fig. 11. Peer synchrony versus peer-hop from the Encoding Server. (a) ARD. (b) Cuatro.

Fig. 12. Packet loss rate versus peer-hops from Encoding Server. (a) ARD. (b) Cuatro.

milliseconds range, attesting to the suitability of the network architecture to carrylive media streaming on large-scale P2P networks. C. Effectiveness of ECC in Isolating Loss Now, we investigate the effects of overlay sizes on the performance scalability of the Zattoo system. Here, we focus on client-side quality (e.g., loss rate). As described in Section I, Zattoo-broadcast media streams are RS-encoded, which allows peers to reconstruct a full stream once they obtain at least of substreams. Since the ECC-encoded stream reconstruction occurs hop by hop in the overlay, it can mask sporadic packet losses and thus prevent packet losses from being propagated throughout the overlay at large. To see if such packet-loss containment actually occurs in the production system, we run the following experiments. We let our monitoring client join a game channel, stay tuned for 15 s, and then leave. We wait for 10 s after the 15-s session is over. We repeat this process throughout the 2-h game coverage and collect the logs. A given log le from each session contains a complete list of packet sequence numbers received from the connected peers during the session, from which we can detect upstream

packet losses. We then associate individual packet loss with the peer-path from the Encoding Server. This gives us a rough sense of whether packets traversing longer hops would experience higher loss rate. In our analysis, we discount packets delivered for the rst 3 s of a session to allow the PDM to stabilize. Fig. 12 shows how the average packet loss rate changes across ). On both channels, different overlay depths (in all cases the packet loss rate does not grow with overlay depth. This result conrms our expectation that ECC helps localize packet losses on the overlay. VI. PEER-TO-PEER SHARING RATIO The premise of a given P2P systems success in providing scalable stream distribution is sufcient bandwidth sharing from participating users [5]. Section IV-A shows that the average sharing ratio of Zattoo users ranges from 0.2 to 0.35, which translates into bandwidth uplinks ranging from 100 to 175 kb/s. This is far lower than the numbers reported as typical uplink bandwidth in countries where Zattoo is available [6]. Aside from the possibility that the users bandwidth may be shared with other applications, we nd factors such as users behavior, support for variable-bit rate encoding, and heterogeneous NAT

66

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

joining peers. This conservativeness necessarily leads to underutilization of resources. C. NAT Reachability Asymmetric reachability imposed by prevalent NAT boxes adds to the difculty in achieving full utilization of the users uplink capacity. Zattoo delivery system supports six different NAT congurations: open host, full cone, IP-restricted, port-restricted, symmetric, and UDP-disabled, listed in increasing degree of restrictiveness. Not every pair-wise communication can occur among different NAT types. For example, peers behind a port-restricted NAT box cannot communicate with those of symmetric NAT type (we documented a NAT reachability matrix in an earlier work [5]). To examine how the varied NAT reachability comes into play as far as sharing performance is concerned, we performed the following two controlled experiments. In both experiments, we run four Zattoo clients, each with a distinct NAT type, tuned to the same channel concurrently for 30 min. In the rst experiment, we xed the maximum allowable uplink capacity (denoted max_cp) of those clients to the same constant value.3 In the second experiment, we let the max_cp of those clients self-adjust over time (which is the default setting of Zattoo player). In both experiments, we monitored how the uplink bandwidth utilization ratio (i.e., ratio between actively used uplink bandwidth and maximum allowable bandwidth max_cp) changes over time. The experiments were repeated three times for reliability. Fig. 14 shows the capacity utilization ratio from these two experiments for four different NAT types. Fig. 14(a) visualizes how fast the capacity utilization ratio converges to 1.0 over time after the client joins the network. Fig. 14(b) plots the average utilization ratio (i.e., ). In both constant and adjustable cases, the results clearly exemplify the varied sharing performance across different NAT types. Especially, the symmetric NAT type, which is the most restrictive among the four, shows inferior sharing ratio compared to the rest. Unfortunately, however, the symmetric NAT type is the second most popular NAT type for Zattoo population [5] and can therefore seriously affect Zattoos overall sharing performance. There is a relatively small number of peers running IP-restricted NAT-type, hence its sharing performance has not been ne-tuned at Zattoo. D. Reliability of NAT Detection To allow communications among clients behind a NAT gateway (i.e., NATed clients), each client in Zattoo performs a NAT detection procedure upon player startup to identify its NAT conguration and advertises it to the rest of the Zattoo network. Zattoos NAT detection procedure implements a UDP-based standard STUN protocol that involves communicating with external STUN servers to discover the presence/type of a NAT gateway [9]. The communication occurs in UDP with no reliable transport guarantee. Lost or delayed STUN UDP packets may lead to inaccurate NAT detection, preventing NATed
3The experiments were performed from one of Zattoos data centers in Europe, and there was sufcient uplink bandwidth available to support 4 * max_cp.

Fig. 13. Receiver-peer churn frequency.

environments contributing to suboptimal sharing performance. Designers of P2P streaming systems must pay attention to these factors to achieve good bandwidth sharing. However, one must also keep in mind that improving bandwidth sharing should not be at the expense of compromised user viewing experience, e.g., due to more frequent uplink bandwidth saturation. A. User Churns It is known that frequent user churns accompanied by short sessions, as well as ash crowd behavior, pose a unique challenge in live streaming systems [7], [8]. User churns can occur in both upstream (e.g., forwarding peers) and downstream (e.g., receiver peers) connectivity on the overlay. While frequent churns of forwarding peers can cause quality problems, frequent changes of receiver peers can lead to underutilization of a forwarding peers uplink capacity. To estimate receiver peers churn behavior, we visualize in Fig. 13 the relationship between session length and the frequency of receiver-peer churns experienced during the sessions. We studied receiver-peer churn behavior for the Cuatro channel by using Zattoos feedback logs described in Table II. The -value at session length min denotes the average number of receiver-peer churns experienced for sessions with length . According to the gure, the receiver-peer churn frequency tends to grow linearly with session length, and peers experience approximately one receiver-peer churn every minute. B. Variable-Bit Rate Streams Variable-bit rate (VBR) encoding is commonly used in production streaming systems, including Zattoo, due to its better quality-to-bandwidth ratio compared to constant-bit rate encoding. However, VBR streams may put additional strain on the peers bandwidth contribution and quality optimization. As described in Section II-C, the presence of VBR streams require peers to perform measurement-based admission control when allocating resources to set up a virtual circuit. To avoid degradation of service due to overloading of the uplink bandwidth, the measurement-based admission control module must be conservative both in its reaction to increases in available bandwidth and its allocation of available bandwidth to newly

CHANG et al.: LIVE STREAMING WITH RECEIVER-BASED PEER-DIVISION MULTIPLEXING

67

Fig. 14. Uplink capacity utilization ratio. (a) Constant maximum capacity. (b) Adjustable maximum capacity.

clients from contacting each other, and therefore adversely affecting their sharing performance. To understand the reliability and accuracy of STUN-based NAT detection procedure, we utilize Zattoos session database (Table I), which stores among other things NAT information, public IP address, and private IP address for each session. We assume that all the sessions associated with the same public/private IP address pair would be generated under the same NAT conguration. If sessions with the same public/private IP address pair report inconsistent NAT detection results, we consider those sessions as having experienced failed NAT detection and apply the majority rule to determine the correct result for that conguration. Fig. 15 plots the daily trend of NAT detection failure rate derived from ARD, Cuatro, and SF2 during the month of June 2008. The NAT detection failure rate at hour is the number of bogus NAT detection results divided by the total number of NAT detections occurring on that hour. The total number of NAT detections indicates how busy Zattoo system was throughout. The NAT detection failure rate grows from around 1.5% at 23 a.m. to a peak of almost 6% at 8 p.m. This means that at the busiest times, the NAT type of about 6% of clients are not determinable, leading to them not contributing any bandwidth to the P2P network. VII. RELATED WORK Aside from Zattoo, several commercial peer-to-peer systems intended for live streaming have been introduced since 2005, notably PPLive, PPStream, SopCast, TVAnts, and UUSee from China and Joost, Livestation, Octoshape, and RawFlow from the EU. A large number of measurement studies have been done on one or the other of these systems [7], [10][16]. Many research prototypes and improvements to existing P2P systems have also been proposed and evaluated [4], [17][24]. Our study is unique in that we are able to collect network core data from a large production system with over 3 million registered users with intimate knowledge of the underlying network architecture and protocol. P2P systems are usually classied as either tree-based push or mesh-based swarming [25]. In tree-based push schemes,

Fig. 15. NAT detection failure rate.

peers organize themselves into multiple distribution trees [26][28]. In mesh-based swarming, peers form a randomly connected mesh, and content is swarmed via dynamically constructed directed acyclic paths [21], [29], [30]. Zattoo is not only a hybrid between these two, similar to the latest version of Coolstreaming [4]; its dependence on Repeater nodes also makes it a hybrid of P2P network and a content-distribution network (CDN), similar to PPlive [7]. VIII. CONCLUSION We have presented a receiver-based, peer-division multiplexing engine to deliver live streaming content on a peer-to-peer network. The same engine can be used to transparently build a hybrid P2P/CDN delivery network by adding Repeater nodes to the network. By analyzing a large amount of usage data collected on the network during one of the largest viewing events in Europe, we have shown that the resulting network can scale to a large number of users and can take good advantage of available uplink bandwidth at peers. We have also shown that error-correcting code and packet retransmission can help improve network stability by isolating packet losses and preventing transient congestion from resulting in PDM recongurations. We have further shown that the PDM and

68

IEEE/ACM TRANSACTIONS ON NETWORKING, VOL. 19, NO. 1, FEBRUARY 2011

adaptive PDM schemes presented have small enough overhead to make our system competitive to digital satellite TV in terms of channel switch time, stream synchronization, and signal lag. ACKNOWLEDGMENT The authors would like to thank A. Goodman for developing the Authentication Server, A. Saurin for the error-correcting code, Z. Wang for an early version of the stream buffer management, E. Wucherer for the Rendezvous Server, Z. Steindler for the Subsidy System, and the rest of the Zattoo development team for fruitful collaboration. REFERENCES
[1] R. auf der Maur, Die Weiterverbreitung von TV- und Radioprogrammenber IP-basierte Netze, in Entertainment Law, F. d. Schweiz, Ed., 1st ed. Bern, Switzerland: Stmpi Verlag, 2006. [2] Euro2008, UEFA [Online]. Available: http://www1.uefa.com/ [3] S. Lin and D. J. Costello Jr., Error Control Coding, 2nd ed. Englewood Cliffs, NJ: Pearson Prentice-Hall, 2004. [4] S. Xie, B. Li, G. Y. Keung, and X. Zhang, CoolStreaming: Design, theory, and practice, IEEE Trans. Multimedia, vol. 9, no. 8, pp. 16611671, Dec. 2007. [5] K. Shami et al., Impacts of peer characteristics on P2PTV networks scalability, in Proc. IEEE INFOCOM, Apr. 2009, pp. 27362740. [6] Bandwidth test statistics across different countries, Bandwidth-Test.net [Online]. Available: http://www.bandwidth-test. net/stats/country/ [7] X. Hei, C. Liang, J. Liang, Y. Liu, and K. W. Ross, Insights into PPLive: A measurement study of a large-scale P2P IPTV system, in Proc. IPTV Workshop, Int. World Wide Web Conf., May 2006. [8] B. Li et al., An empirical study of ash crowd dynamics in a P2Pbased live video streaming system, in Proc. IEEE GLOBECOM, 2008, pp. 15. [9] J. Rosenberg et al., STUNSimple traversal of User Datagram Protocol (UDP) through network address translators (NATs), RFC 3489, 1993. [10] A. Ali, A. Mathur, and H. Zhang, Measurement of commercial peer-to-peer live video streaming, in Proc. Workshop Recent Adv. Peer-to-Peer Streaming, Aug. 2006. [11] L. Vu et al., Measurement and modeling of a large-scale overlay for multimedia streaming, in Proc. Int. Conf. Heterogeneous Netw. Quality, Rel., Security Robustness, Aug. 2007, pp. 17. [12] T. Silverston and O. Fourmaux, Measuring P2P IPTV systems, in Proc. ACM NOSSDAV, Jun. 2007. [13] C. Wu, B. Li, and S. Zhao, Magellan: Charting large-scale peer-topeer live streaming topologies, in Proc. ICDCS, 2007, p. 62. [14] M. Cha et al., On next-generation telco-managed P2P TV architectures, in Proc. Int. Workshop Peer-to-Peer Syst., Feb. 2008. [15] D. Ciullo et al., Understanding P2P-TV systems through real measurements, in Proc. IEEE GLOBECOM, Nov. 2008, pp. 16. [16] E. Alessandria et al., P2P-TV systems under adverse network conditions: A measurement study, in Proc. IEEE INFOCOM, Apr. 2009, pp. 100108. [17] S. Banerjee et al., Construction of an efcient overlay multicast infrastructure for real-time applications, in Proc. IEEE INFOCOM, 2003, vol. 2, pp. 15211531. [18] D. Tran, K. Hua, and T. Do, ZIGZAG: An efcient peer-to-peer scheme for media streaming, in Proc. IEEE INFOCOM, 2003, vol. 2, pp. 12831292. [19] D. Kostic et al., Bullet: High bandwidth data dissemination using an overlay mesh, in Proc. ACM SOSP, Bolton Landing, NY, Oct. 2003, pp. 282297. [20] R. Rejaie and S. Stafford, A framework for architecting peer-to-peer receiver-driven overlays, in Proc. ACM NOSSDAV, 2004, pp. 4247.

[21] X. Liao et al., AnySee: Peer-to-peer live streaming, in Proc. IEEE INFOCOM, Apr. 2006, pp. 110. [22] F. Pianese et al., PULSE: An Adaptive, incentive-based, unstructured P2P live streaming system, IEEE Trans. Multimedia, vol. 9, no. 8, pp. 16451660, Dec. 2007. [23] J. Douceur, J. Lorch, and T. Moscibroda, Maximizing total upload in latency-sensitive P2P applications, in Proc. ACM SPAA, 2007, pp. 270279. [24] Y.-W. Sung, M. Bishop, and S. Rao, Enabling contribution awareness in an overlay broadcasting system, in Proc. ACM SIGCOMM, 2006, pp. 411422. [25] N. Magharei, R. Rejaie, and Y. Guo, Mesh or multiple-tree: A comparative study of live P2P streaming approaches, in Proc. IEEE INFOCOM, May 2007, pp. 14241432. [26] V. N. Padmanabhan, H. J. Wang, and P. A. Chou, Resilient peer-topeer streaming, in Proc. IEEE ICNP, Nov. 2003, pp. 1627. [27] M. Castro et al., SplitStream: High-bandwidth multicast in cooperative environments, in Proc. ACM SOSP, Oct. 2003, pp. 298313. [28] J. Liang and K. Nahrstedt, DagStream: Locality aware and failure resilient peer-to-peer streaming, in Proc. SPIE Multimedia Comput. Netw., Jan. 2006, vol. 6071. [29] N. Magharei and R. Rejaie, PRIME: Peer-to-peer receiver-drIven MEsh-based streaming, in Proc. IEEE INFOCOM, May 2007, pp. 14151423. [30] M. Hefeeda et al., PROMISE: Peer-to-peer media streaming using CollectCast, in Proc. ACM Multimedia, Nov. 2003, pp. 4554. Hyunseok Chang received the B.S. degree in electrical engineering from Seoul National University, Seoul, Korea, in 1998, and the M.S.E. and Ph.D. degrees in computer science and engineering from the University of Michigan, Ann Arbor, in 2001 and 2006, respectively. He is currently a Member of Technical Staff with the Network Protocols and Systems Department, Alcatel-Lucent Bell Labs, Holmdel, NJ. Prior to that, he was a Software Architect at Zattoo, Inc., Ann Arbor, MI. His research interests include content distribution network, mobile applications, and network measurements.

Sugih Jamin received the Ph.D. degree in computer science from the University of Southern California, Los Angeles, in 1996. He is an Associate Professor with the Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor. He cofounded a peer-to-peer live streaming company, Zattoo, Inc., Ann Arbor, MI, in 2005. Dr. Jamin received the National Science Foundation (NSF) CAREER Award in 1998, the Presidential Early Career Award for Scientists and Engineers (PECASE) in 1999, and the Alfred P. Sloan Research Fellowship in 2001.

Wenjie Wang received the Ph.D. degree in computer science and engineering from the University of Michigan, Ann Arbor, in 2006. He is a Research Staff Member with IBM Research China, Beijing, China. He cofounded a peer-to-peer live streaming company, Zattoo, Inc., Ann Arbor, MI, in 2005, based on his Ph.D. dissertation, and was CTO. He also worked with Microsoft Research China and Oracle China as a visiting student and consultant intern.