1 Insights on Media Streaming Progress using BitTorrent-like Protocols for On-Demand Streaming Nadim Parvez, Carey Williamson, Anirban Mahanti, and Niklas Carlsson Abstract—This paper develops analytic models that charac- media case has greater temporal diversity of requests. The terize the behavior of on-demand stored media content delivery peer dynamics resemble those of file downloading, while still using BitTorrent-like protocols. The models capture the effects requiring low startup delays for the sequential playback of of different piece selection policies, including Rarest-First, two variants of In-Order, and two probabilistic policies (Portion and large media objects. Finally, live streaming implicitly involves Zipf). Our models provide insight into system behavior, and sustained content delivery at the intrinsic media playback rate, help explain the sluggishness of the system with strict In-Order while the stored media case is general: the retrieval rate could streaming. We use the models to compare different retrieval be slower than, faster than, or the same as the media playback policies across a wide range of system parameters, including peer rate, or even vary with time. arrival rate, upload/download bandwidth, and seed residence time. We also provide quantitative results on the startup delays These characteristics can challenge the performance of and retrieval times for streaming media delivery. Our results existing P2P protocols. For example, BitTorrent improves the provide insights into the design tradeoffs for on-demand media efficiency of file downloads by using a Rarest-First piece streaming in peer-to-peer networks. Finally, the models are selection policy to increase the diversity of pieces available validated using both fluid-based and packet-level simulations. in the network. However, streaming protocols require in-order Index Terms—Peer-to-peer systems, BitTorrent, On-demand playback of media content, which naturally implies that in- streaming, Modeling order retrieval of pieces is desirable (but not strictly required). In-order collection of pieces may reduce the spatial and I. INTRODUCTION temporal diversity of pieces in a P2P network, resulting in poor system performance. Peer-to-peer (P2P) networks offer a promising approach for In this paper, we analytically characterize the performance Internet-based media streaming. P2P networks are autonomous of BitTorrent-like protocols for on-demand streaming of stored systems with the advantages of self-organization and self- media files. Our models capture performance differences be- adaptation. P2P solutions can enable efficient and scalable tween various policies and configuration details (e.g., piece media streaming, as long as they can meet the sequential selection policies, upload bandwidth) and allow us to answer playback demands of media streaming applications, which questions related to the efficency and user-perceived perfor- differ from those of file downloading, for which P2P file mance of BitTorrent-like on-demand streaming protocols. sharing networks were originally created. The main contributions in our paper are the following: The P2P paradigm has been used successfully for live • media streaming, but the (more difficult) case of on-demand We show that the analysis of P2P media streaming is streaming of stored media has received relatively less attention. decomposable into download progress and sequential The two scenarios share several common challenges, including progress, which can be analyzed separately. Furthermore, the sequential playback demands of large media objects, improving one component can usually be done without the geographic diversity of heterogeneous receivers, and the compromising the other. • dynamic churn of the media streaming population. We develop detailed analytical models that explicitly On-demand streaming of stored media files differs in subtle consider piece selection policies. The models accurately but important ways from live media streaming. First, live predict the transition rate of downloaders to seeds, as well streaming typically involves only a single streaming source, as the steady-state swarm population size and mix. The whereas stored media streaming can involve many providers of models provide important insights into the efficiency of content. Second, the stored media case involves retrieving the on-demand media streaming in P2P networks. • entire media object, while the live streaming case allows peers The models explicitly consider the number of upload and to join at any time (i.e., mid-stream), without retrieving earlier download connections, rather than just the total network portions of the stream. Thus the issue of “startup delay” differs bandwidth [26], [29]. This formulation provides the flex- in the two scenarios (i.e., joining an existing stream versus ibility to model concurrent connections and consider the starting a new stream). Third, the peers in a live streaming effect of network asymmetry on the system performance. • scenario have a shared temporal content focus, while the stored The models provide estimates of the expected retrieval time for stored media objects, as well as its variability, so C. Williamson is with the Department of Computer Science, Uni- that we can determine suitable tradeoffs between startup versity of Calgary, Canada. Email: [email protected] A. Mahanti delay and the likelihood of uninterrupted streaming. is with National ICT Australia (NICTA), Eveleigh, NSW, Australia. Email:[email protected] N. Carlsson is with Link¨oping Univer- • The models are validated using both fluid-based and sity, Link¨oping, Sweden. Email:[email protected] packet-level simulations. 2 The remainder of the paper is organized as follows. Sec- This section introduces the concept of media streaming tion II presents a brief description of the BitTorrent system, progress (MSP), which is defined as the number of useful as well as the concepts of download progress and sequen- media pieces obtained per unit time. Conceptually, the MSP tial progress. Sections III and IV explain the derivation of can be separated into two parts: (1) the download progress basic models for sequential progress and download progress, (DP), which is defined as the number of pieces retrieved respectively, using different piece selection policies. Section V per unit time, and (2) the sequential progress (SP), which is presents the analysis of startup delay, and also discusses sev- defined as the number of useful in-order media pieces obtained eral extensions of our model. Section VI presents simulation per piece retrieved. The MSP is simply the product of these results to validate the models. Section VII summarizes relevant two metrics. Equation 1 expresses this simple relationship. related work, while Section VIII concludes the paper. MSP = DP × SP (1) II. BACKGROUND AND SYSTEM TRADEOFF The download progress captures the generic notion of A. BitTorrent throughput (i.e., a policy’s ability to download pieces quickly), BitTorrent [9] is a popular peer-to-peer file sharing system while the sequential progress refers to an application-specific used to facilitate efficient downloads. BitTorrent splits files property, namely the ability of a piece selection policy to into pieces, which can be downloaded in parallel from different acquire the initial pieces from the beginning of a file, as peers. BitTorrent distinguishes between peers that have the required for streaming media playback. Note that sequential entire file (called seeds), and peers that only have parts of the progress (the sequentiality of the pieces obtained) is conceptu- file (called leechers or downloaders) and are still downloading ally independent of the download progress (the rate at which the rest of it. The set of peers collaborating to distribute a the pieces are obtained). In the following sections, we analyze particular file is known as a BitTorrent swarm. these metrics separately, starting with sequential progress. A tracker maintains information about the peers partici- pating in a swarm. New peers wanting to download a file are directed (typically using information provided in a meta- III. SEQUENTIAL PROGRESS file available at an ordinary Web server) to a tracker, which In this section we analyze the sequential progress of four provides each new peer with the identity of a random set of simple policies: strict In-Order retrieval, Random piece selec- participating peers. Each peer typically establishes persistent tion, and two probabilistic piece selection policies (Portion connections with a large set of peers, consisting of peers and Zipf [3]). An example of sequential progress for each is identified by the tracker as well as by other peers to which shown in Figure 1. the peer is connected. The peer maintains detailed information By definition, the strict In-Order policy is ideal in terms of about which pieces the other peers have. sequential progress. Each peer simply retrieves the file pieces While peers can typically request pieces from all connected in numerical order from 1 to M. However, the download peers that have useful pieces, each peer only uploads to a progress of this policy in a P2P network can be sluggish, as limited number of peers at any given time. That is, most will be seen in Section IV-D. peers are choked, while a few peers that it is currently The Random piece selection policy provides poor sequential willing to serve are unchoked. To encourage peers to upload progress, as shown in Figure 1. While not the worst case1 pieces, BitTorrent