Multimedia transmission in a managed P2P network: making sense?

Multimedia transmission in a managed P2P network: making sense? L. Xu 1, A. Ksentini 2, K. Singh 1, G. Rubino 1, G. Straub 3, Y. L Azou 4 1 INRIA Rennes - Bretagne Atlantique, Rennes, France; 2 IRISA-University of Rennes 1, Rennes, France; 3 Thomson, Rennes, France, 4 France Telecom, Lannion, France E-mail: Letian.Xu@irisa.fr; adlen.ksentini@irisa.fr, kamal.singh@irisa.fr, rubino@irisa.fr, gilles.straub@thomson.net, yves.lazou@orange-ftgroup.com Abstract: The peer-to-peer (P2P) paradigm has recently gained tremendous attraction and is widely used for content distribution and file sharing. In this context, digital media companies have recently started embracing peer-assisted distribution networks as an alternative to traditional Client/Server architectures. P2P systems can ensure a fast and scalable delivery of media content. However they suffer from the total absence of control on content distribution as they are running on PCs and are managed by users themselves. Users PCs are unsecure, unmonitored, and almost impossible to control by a service or network provider. One solution to this problem is to use Managed P2P networks, where devices included in the P2P networks are controlled by the provider. In this paper, we illustrate the benefit from using such a Managed P2P system in the context of VoD distribution within an IPTV network. For this purpose, we compare the performance of VoD service based on the classical Client/Server model and a Managed P2P solution. The results obtained can be considered as a first insight on the interest of deploying P2P architectures inside managed networks. Keywords: P2P, Managed Networks, VoD, Multimedia,CDN 1 INTRODUCTION Multimedia services are expected to be a significant source of revenue for service providers as their demand is rapidly increasing with the explosive growth of Internet. Video on Demand (VoD) services are among the most popular multimedia applications. They allow users to watch any movie (in some data base) at any given time. Observe that a high bandwidth is required to watch a good quality movie. When the total number of clients increases, the total number of required VoD servers, that are costly devices, also increases. This makes classical Client/Server solutions cost-ineffective. As the demand grows, more servers must be added to the system, which means more cost, changes, bring also more costs on maintenance, etc. Peer-to-peer (P2P) systems can be a solution to the above problem because they inherently have the property of scalability. In a P2P network, each client acts as a server and other clients can download data from him. This means that when the number of clients increases, then the number of available servers, that are actually also clients acting as servers, also increases, without necessarily changing the number of actual VoD servers. Moreover, clients (called also peers in this context), can also utilize the unused upload bandwidth of other clients acting as servers. This decreases the load on the VoD servers and thus, theoretically at least, with P2P systems the requirements in terms of VoD servers and bandwidth capacity of service providers are reduced. This can significantly diminish the deployment cost of these VoD services, which in turn mean low costs for end users and possibly an increase of their overall popularity. Nevertheless, the efficient design of a P2P system for VoD is not a trivial task, given the constraints to satisfy in order to operate with efficiency enough and to achieve those theoretical gains. At present, video services are distributed either over public networks or over private networks that we call here managed networks. In the latter, unlike public network that offers only Best-Effort service, the service provider can control the system in many ways. Furthermore, multicast routing is often enabled on private networks for live TV transmission. For instance, when telecom operators provide triple play services, they deliver video services over private managed networks using multicast IP. If the interest for P2P on public networks is obvious for live TV transmission (due to the absence of multicast routing on the public network), it may be less evident in the case of a managed network. Therefore the interest of P2P for managed networks can only reside in unicast services such as in VoD, by decreasing the deployment cost. Observe that we are addressing here only the basic service, movies distribution; for more intelligent services (e.g., time shifting) other factors may make P2P systems very attractive. In this paper we discuss about basic benefits from using P2P ideas over Managed networks in the context of VoD distribution. We compare the performance of a VoD service when using the classical Client/Service and a Managed P2P version. This paper is organized as follows. Section 2 gives some background on P2P architectures as well as a brief description on their use for multimedia applications. Section 3 introduces the concept of managed P2P networks. Section 4 compares the performances of VoD distribution when using Managed P2P systems and Client/Server architectures. We conclude the paper in Section 5. 2 MULTIMEDIA SERVICES OVER P2P Multimedia services using the P2P paradigm are gaining popularity over the Internet. Not only VoD, but live streaming Corresponding author: Ksentini, IRISA-University of Rennes 1, Campus Universitaire de Beaulieu, Rennes (France), 02 99 84 71 42, adlen.ksentini@irisa.fr

services like IPTV are using P2P today. We describe some of these applications in this section. Figure 1. Single-Source P2P streaming model For providing multimedia content, two types of network architectures are usually possible. A single source architecture (see Figure 1) means that each peer downloads the multimedia content from only one source. Of course these downloading peers can also act as sources for other peers, but no more than one source will be proving content to a given client. A single source architecture was initially proposed by Peercast [1]. Peercast was created in 2002 and is the first application to allow users to broadcast their own content using P2P. PeerTV [2], created in 2006, also uses a single-source architecture and relies on the Peercast protocol. Figure 2. Multiple-Source P2P streaming model Multiple sources architectures (see Figure 2) are used when several source peers offer the content to one peer. In this case, the content is divided into several pieces and these pieces are forwarded in the network. In other words, a peer gets the multimedia content from different peers and sends itin turn to one or many client peers. This kind of architecture is used for instance by PPLive [3] and other applications described below. PPLive is a proprietary application already deployed at a large scale, and is today the most popular live TV system over the Internet. Some information about the way it works has been obtained through measurements and by looking at the PPLive messages. It encodes the videos and divides it into chunks. The peers upload and download different chunks from other peers. In order to ensure a smooth play out of the video, PPLive uses a combination of two buffers that store the incoming data in order to handle bandwidth variations in the Internet. In spite of this, play-out interruptions can not always avoided because of those frequent changes in the available bandwidth Scribe and SplitStream are both application-layer group communication systems developed by Microsoft Research. SplitStream, as indicated in its name, splits a multicast video stream into many stripes, and transmits them in a forest like multicast path. Joost [4] is the new P2P application by the creators of Kazaa and Skype, formerly known as the Venice project. It is a P2P video streaming application allowing accessing to TV channels for free, even if up to now, the number of channels is limited and the broadcasted content is not live. But this may change because Joost is signing agreements with major broadcasting companies. Joost is a streaming application only. Users can not download or store content on their hard-disk and is a P2P application because clients can get data from other peers. However, some fixed nodes are also present in the network architecture. Joost works then like a CDN (Content Distribution Networks) network augmented with P2P facilities when the content is popular and well distributed. PULSE is a peer-to-peer live streaming system designed to operate in scenarios where the nodes bandwidth resources can be highly heterogeneous and variable over time, such as today s Internet. To support such real-life network conditions, one has to carefully think over the quality vs. timeliness tradeoff. In the past, the system designer s attention has been primarily focused on timeliness. Striving to obtain the lowest possible latencies often resulted in simple structured system designs, where tree-based overlays were built following some optimizing criteria: since trees scale well with respect to maximum path lengths and node placements, they can be optimized with respect to bandwidth or latency. Media quality, however, suffers when node instability is present, as it happens when there are rapid variations in the population due to node arrivals or departures. 3 FROM UNMANAGED P2P TO MANAGED P2P One problem with the P2P model is the total absence of control on content distribution as it is managed by users themselves. Users devices are unsecure, unmonitored, and almost impossible to control by a service or network provider. With P2P, traditional telecom operators see their role relegated to Internet connectivity providers and are left out of the loop of content distribution business. On the other hand, content owners (such as TV broadcasters, Movie studios, game designers) see in P2P a content distribution solution tinted with a piracy history, and not able to ensure the service level guarantees they demand. In general P2P applications suffer from a number of problems: - Very inefficient use of network resources (the overlay built by the peers is not optimized taking into account the underlay, i.e. the actual network), and consequently very poor performance. The uplink access bandwidth dramatically impacts uploading performance because of the P2P incentive model to share content.

- Lack of service guarantees due to uncontrolled interference between the different applications, which often results in poor quality for P2P-based applications such as live video streaming or even telephony. - Absence of security and control that makes it impossible to guarantee the integrity and security of the content, which limits the quality and the diversity of available content. With respect to Content Distribution, P2P has probably been mostly successful so far in terms of data volume distributed. This lies in the fact that the content distribution efforts are shared across all computers at the edge of the network, and that it does not require the deployment of expensive and complex servers. However, the success of P2P has also been the cause of most of its difficulties. Current P2P systems operate over non-managed, non-trusted end-user PCs, which have produced a myriad of security, piracy, privacy, and poor end-user experience problems. In particular: (i) P2P performance is unpredictable. Content is not easily available where it is mostly needed, no SLAs can be provided to content providers, user links are congested as P2P downloaders are encouraged to serve many people to get access to more content, etc; (ii) P2P systems became popular carrying illegal content. As a consequence, most people associate with peerto-peer traffic the idea of illegal traffic. This language abuse makes ISPs and content providers reluctant to adopt this content distribution technology; (iii) It is happening in a noncohesive and non-standardised way; (iv) ISPs do not participate in the P2P cycle and cannot monetize the usage of P2P services. P2P networks are not optimised to be ISP friendly. All this can change if industry decides to build on P2P technology and to use it as a new content distribution paradigm. Defining a managed P2P system by implementing a fully distributed solution on devices controlled by a telecom operator will bring the following advantages: - Core technology runs on managed devices now controlled by a carrier or a virtual ISP. - Always On devices and stable topologies can be thus available. - Resources can be allocated in a distributed manner without worrying about preventing free-riders, etc. - A controlled platform reduces piracy risks. - A controlled infrastructure leads to lower costs and increased network robustness. - ISPs can be allowed to enter the content distribution loop, offering better QoS, billing, managed services, etc. - Better understanding and exploitation of user preferences: such information is used to better position content (including publicity). - Allows ISPs to do troubleshooting. The Internet Gateway is the best place to host a peer to peer client, because it is always on and under the control of the operator. It is not an open platform (as the PC is), therefore reducing piracy risks is feasible. In the sequel, we show that using P2P networks for other situations than distributing illegal content is possible. Rather than being only Internet connectivity provider, network provider can thus participate to content distribution business by proposing Multimedia services over Managed P2P. For instance, P2P can be used in managed networks (also called IPTV networks) to distribute unicast IPTV services such as VoD, or catch-up TV in a secure manner, with low cost and high performances. Catch-up TV consists in making available in a VoD server some TV programs that have been already broadcasted. When a user missed a dedicated show because he was not at home, he can get it later on using this catch-up TV service. Catch-up TV is also a unicast service that if often made available both on Open Internet and on managed networks. In the rest of the document we restrict ourselves to the VoD application even if this is applicable to other unicast IPTV services. 4 VOD DISTRIBUTION: MANAGED P2P VERSUS CLIENT/SERVER In this section, we compare the performance of VOD services when using Managed P2P and Unicast transmission. By Unicast transmission, we mean a classical Client/Server architecture. To capture the performances of VOD service when using Unicast we draw very simple mathematical models. For the Managed P2P however, we use simulation results by means of ns2 (Network Simulator 2) tool [5]. 4.1 VOD distribution with unicast transmission To draw the performance of the Unicast model we use the following notation and simplifying assumptions: - N is the number of clients, - B is the bandwidth of the source, in Mbps, - D is the download bandwidth of the customers, in Mbps, assumed to be the same for all clients. Using UDP transmission, the number of clients served by one VOD server is simply: where Int(x) denotes the integer part of x. Here, we consider only server s network capacity; no attention will be given to the computational capacity of the VOD server. The time T spent by a VOD user to download the whole file to all the clients (simultaneously) is Based on (1) and (2) we draw the performance of VOD when using the classical Client/Server architecture (see below). (1) (2)

4.2 VOD distribution with Managed P2P Concerning the performance of a VoD service in Managed P2P, we used the BitTorrent protocol. This protocol [6] is proposed with the objective to disseminate one large file (or a composition of large files) to a large number of users in an efficient way. Therefore, for each file an overlay network is created. The file sharing is based on the swarming principle, which is also denoted as multi-source download. The file of interest is fragmented into chunks. When a peer completes the download of a single chunk, it offers it to other peers which so far have not downloaded it. Thus, peers exchange chunks with each other although they did not finish downloading the complete file. Therefore, the resources in the P2P network are used more efficiently and the network also scales for large peer populations with respect to download time. According to the original BitTorrent specification, each overlay network consists of two different kinds of peers, the seeds and the leechers, and a so-called tracker. A seed holds the complete file and uploads it to others altruistically, whereas a leecher is still downloading the file. The tracker is a centralized component which stores information about all peers. A new peer, which enters the network, asks the tracker for a list of active peers in the overlay. The tracker returns a random subset to the requesting peers. Furthermore, an active peer contacts the tracker from time to time to obtain information about new peers in the network. An extension of the protocol also incorporates the exchange of information about other peers in the network between connected peers. This is often stated as trackerless BitTorrent. In the BitTorent model, we assume that a VoD server is represented by one seeder, and the managed clients is represented by leechers. Since we are in a Managed P2P network, the clients can be forced to stay in the overlay network for serving other peers despite the fact that they finished downloading the required file. For our simulation, we used the ns2 BitTorrent model available in [7]. 4.3 Network model and parameters The network model and parameters were provided by France Telecom * using some available data. Generaly speaking, internet access network is divided in several parts, see [8]: Customer Premises Network, Access & Aggregation Network, and Regional Broadband Network. Over this topology, PPP protocol is widely used as a way to access internet: this protocol conveys client's packets towards a termination point in a BRAS (Broadband Remote Access Server). BRAS equipment is often located with VoD servers in a place called PoP (Point Of Presence), at the edge of Access & Aggregation and Regional networks. Several PoPs are cleverly spread in the network. Based on the assumption that the bottleneck of the network is at the access links of the users (typically ADSL loop line) and not at the routers in the rest of the network, we simplify the topology as a star model shown in Figure 3. The central Router stands for the BRAS and the PoP location. * French Service/Network Provider The bandwidth B of the source is B = 100 Mbps. The clients are connected to a router through an asymmetric link, where the download bandwidth D is equal to 2 Mbps, which is used by a wide range of ADSL clients. The size of the file is equal to 100 Mb. For exploring the P2P network s behaviour, we developed three simulation models. In the first model, the P2P architecture operates at the TCP-level; the TCP behaviour is thus taken into account in the simulations. The uploading bandwidth of each peer application, including the seeds, is uniformly distributed on a range going from 64 kbps to 1024 kbps. These values are based on BitTorrent software, where users can limit it upload capacity. The second model for the simulations neglects the TCP behaviours and considers only the application-layer (peer layer). It also fixes the uploading capacity of each peer, which is much closer to the situation that can be encountered in a managed network. While in the TCP-level simulations, all the messages are handled by the TCP agents, in the simulations of this second model, the control messages are delivered directly to the receiver whereas the PIECE messages are queued in an upload queue. These PIECE messages are specific to the BitTorrent protocol, and they are used to inform the other peers that a piece of file is proposed to download. So, in this model, all the peers use the managed uploading capacity of 0.5 Mbps, and the seeds use the managed uploading capacity of 100 Mbps, for example, which makes the downloading time less than that of TCP-level simulations. Figure 3. Network Model The third model simulates the leaving and joining behaviours of the peers in a P2P network. In previous simulations, all the peers were fixed and not allowed to leave until the last one finished its downloading. In this third case we added a module to the simulator, in order to control the peers activities. Through this controller, we were able to control the percentage of the peers who leave and rejoin the P2P overlay, and the active mean time (AMT) and inactive mean time (IMT). The active mean time is the mean duration of time interval during which a peer stays connected. The inactive

mean time is the corresponding average duration of the disconnection periods. For the Unicast model we used equations (1) and (2). 4.4 Results Our target was to measure the global downloading time T in which all the peers finish downloading. This is a good metric for evaluating the scalability of both architectures (Managed P2P and Unicast). performance doesn t deteriorate when the number of peers increase, which shows the good scalability of the P2P networks. Note the presence of a crossing point between the Client/Server version and the P2P one, showing that Unicast performs better for low values of N. We argue this behaviour by the fact that: (i) Unicast-based model uses UDP, which represents its best-case situation; (ii) TCP dynamics decreases the performance of P2P as compared with UDP used by Unicast. 70000 60000 Unicast with UDP (theoretical) 100 Peers 200 Peers 300 Peers Download Time (s) with randomness model 50000 40000 30000 20000 10000 0 0 100 200 300 400 500 Number of Peers Figure 4. Download Time (T) of P2P compared to Unicast (Client/Server) Figure 4 presents the download time of the first simulated P2P model compared with that obtained with Unicast. Figure 5. Download Time (T) of application-level compared with that of Unicast In this P2P simulation model, peers never leave the managed network. Different simulations were done. For each run we increased the number of peers (clients) N, while using the same file size (100 MB). For a point m in the x-axis, the corresponding point T(m) in the y-axis is the download time of the m th client to get the file. It is clearly seen that P2P Figure 6. Download Time (T) of application-level P2P compared to that of the Unicast (Client/Server) version In the second model, we assumed that the routing is done at the application level, and that the normal peers uploading capacity is fixed at 0.5 Mbps, while the seeds uploading capacity is fixed at 100 Mbps. The obtained results are shown in Figure 5. Compared to the first model, we can observe that when the number of peers is less than 150, the peers have almost the same downloading time as with Unicast; for larger populations, the P2P version achieves much better scalability, illustrating again how P2P-based VoD transmission achieves better results than an Unicast-based solution. The third model s aim is to study P2P s dynamics, especially with respect to leaving and joining behaviour. As an input, we have 1500 seconds of active mean time, 50 seconds of inactive mean time and 10% of the peers having this behaviour. This dynamics is studied using the TCP-level simulations that have uniform distribution for uploading bandwidths. The results are shown in Figure 6. The peak in the curves corresponds to the downloading time of those clients who left before finishing and then came back. Most of clients have a downloading time with low variability with values similar to the previous ones. This model of simulation corresponds to a non-managed P2P network, where the clients are not controlled. Thus, compared to Figure 4 which shows a completely Managed P2P, we notice that besides ensuring a Of course, this is just an illustration

secure P2P system, managed P2P networks can achieve better results than non-managed P2P and Unicast ones. 5 CONCLUSION In this paper, we investigated the performance of transmitting multimedia Unicast services, particularly VoD (or catch-up TV), over managed P2P networks. By Managed P2P, we meant P2P network where devices (peers) are under Service Provider control. The BitTorrent protocol was used to implement the managed P2P architecture in the simulations, and this protocol was compared to Client/server architecture for distributing VoD services. The obtained results clearly show the gain achieved when using managed P2P over Unicast transmission in terms of download time and system scalability, where managed P2P performance remains independent from the amount of concurrent demand even for huge number. Thus, a natural response to the question P2P in a managed network: making sense? is yes, but only when there are many clients in the system. Acknowledgement This Work has been done within the P2Pim@ges project, funded by the DGE and the Pôle de compétivité Images et Réseaux working in the French Brittany area. References [1] Peercast, http//www.peercast.org. [2] PeerLive, http:// www.peertv.eu. [3] PPLive, http//www.pplive.com [4] Joost, http://www.joost.com. [5] Network Simulator 2, http://www.isi.edu/nsnam/ns. [6] Bittorrent Protocol Specification v1.0. http://wiki.theory.org/bittorrentspecification. [7] Bitorent module for ns2, http://www.tuharburg.de/et6/research/bittorrentsim/index.html [8] DSL Forum TR-101 "Migration to Ethernet-Based DSL Aggregation", April 2006