Impacts of Peer Churn on P2P Streaming Networks* Xiaohan Kang 1 , Juan Jos´ e Jaramillo 2 and Lei Ying 1 Abstract— Peer-to-peer (P2P) technology has been broadly adopted in live media streaming in recent years. In this paper, we consider a P2P streaming network where a server generates content chunks, and transmits each chunk to a randomly selected peer. Peers then exchange chunks among themselves according to some chunk selection policy. While the performance of different chunk selection policies has been intensively analyzed assuming no peer arrival or departure, the impact of peer churn on P2P live streaming has not been well understood yet. We show that unlike in the static network scenario, larger buffer size does not necessarily result in higher playout probability in the presence of peer churn. We further analyze the relation between the buffer size, playout probability and peer churn under different chunk selection policies and characterize the impact of peer churn on the performance of these chunk selection policies via both theoretical analysis and simulations. I. INTRODUCTION Internet video became the largest traffic type in 2010, accounting for 40% of all global consumer traffic, and it is expected to reach 62% by the end of 2015 [1]. Thus, the importance of studying efficient and scalable methods to deliver real-time media is paramount. Because of its inherent scalability, P2P technology has been extremely successful in delivering streaming videos to millions of users. One approach of P2P streaming is to use multicast net- working at the IP or higher layer, where data is pushed along pre-constructed multicast trees [2], [3], [4]. This approach, however, requires significant infrastructural overheads, par- ticularly in networks with peer churn because the multicast trees need to be reconfigured every time a peer arrives or de- parts [5]. An alternative approach is unstructured P2P stream- ing [6], [7], [8]. In these networks, the streaming content is divided into small chunks that are disseminated to a small number of peers, and the content is then duplicated within peers in a gossip-like fashion without the intervention of any centralized servers, so unstructured P2P streaming is robust under peer churn. Because of its robustness, unstructured P2P streaming has become the dominant architecture of large- scale P2P streaming networks, such as CoolStreaming [6], PPTV [9] and PPStream [10]. In an unstructured streaming network, each peer maintains a playout buffer. The chunk at the tail of the buffer is the one to be played out in the current time slot. At the end of each time slot, all the chunks are *Research supported by NSF CNS-1261429. 1 X. Kang and L. Ying are with the School of Electrical, Computer and Energy Engineering, Arizona State University, Tempe, AZ 85287, USA {xiaohan.kang,lei.ying.2}@asu.edu 2 J. J. Jaramillo is with the Department of Applied Math and Engineering, Universidad EAFIT, Medell´ ın, Colombia jjaram93@eafit.edu.co moved one position closer to the tail. In each time slot, every peer randomly contacts another peer to download a chunk missing in the buffer so that when the chunk needs to be played out, it is present in the buffer with high probability. The probability that any peer is able to obtain a specific chunk by the time of playout is called playout probability. Among other metrics, the playout probability and the buffer size are essential to evaluate the performance of a P2P live streaming network since the playout probability measures the continuity of the streaming experience, while the buffer size measures the delay of the playback of the live event. There are two main problems that need to be studied in P2P streaming protocol design: how to select a peer to contact [11] and how to select a chunk to download [12], [13], [14]. Among various chunk selection policies, the rarest- first policy, which tries to obtain the most recently generated chunk, is most popular, while the greedy policy, which attempts to download the chunk closest to playout, is an intuitive alternative. The combination of the two, called the hybrid policy, achieves order sense optimality in minimizing the buffer size when peers are static [15]. In this paper we study the problem of chunk selection, and we focus on the three selection policies described above. The main advantage of unstructured P2P streaming over structured P2P streaming is that the unstructured approach has robust performance under peer churn. However, existing works of unstructured P2P streaming (see [15], [16] and references within) almost exclusively ignore peer churn and assume no peer arrival/departure. This is mainly due to the difficulty of establishing a tractable model of P2P streaming networks under peer churn. To the best of our knowledge there have only been empirical studies of peer churn in P2P file sharing networks such as Gnutella or BitTorrent [17], models for structured P2P streaming as in [18], as opposed to unstructured networks, or via simulations as in [5]. In this paper we study peer churn with a constant arrival and departure rate (churn rate). We try to answer the follow- ing questions in the presence of peer churn: 1) Does larger buffer guarantee higher playout probability? 2) Is there a limit of the playout probability one can achieve? 3) Do some chunk selection policies perform better than others in terms of streaming continuity and delay, and why? We tackle these questions by fluid and discrete methods and simulations. Our contributions are summarized as follows: 1) Contrary to the previous result for the scenario without peer churn, where increasing the buffer size always leads to higher playout probability, we derive theoret- ical upper and lower bounds of the buffer size for the rarest-first policy for a given target playout probability,