On scalability of proximity-aware peer-to-peer streaming☆
Introduction
Live multimedia streaming is gaining increasing popularity with the advent of commercial deployment from major content providers. Among the existing systems, P2P streaming has emerged to be a promising approach to large-scale multimedia distribution [6], [17], [4], [19], [10], [8], [23], [15], [14]. The basic idea is that each peer in the P2P streaming system will contribute their uploading capacities to the downloading demands of other peers. In this way, the media server load could be significantly reduced. Therefore the system is able to support a larger number of peers in a streaming session with a fixed server capacity, and thus achieve better scalability.
While proven to provide a better scalability in terms of server load, the overall P2P streaming system performance in terms of delay and network bandwidth cost largely depends on the P2P topology. At the core of its construction is the problem of peer selection – how to select the parent peer(s) to download the stream. The goal is to construct a high quality topology that could minimize the server load and delay experienced by peers, and reduce the network bandwidth cost. To address this problem, the existing approaches have resorted to intuitions and heuristics. The proposed solutions include tree-based, mesh-based, directed-acyclic-graph-based, and randomized topology construction algorithms and protocols.
One of the fundamental challenges that all these approaches face is the problem of topology mismatch between the overlay layer of P2P network and physical layer network. The peers which are one hop away in the P2P topology could have certain distance (number of hops) between them in the underlying physical network. Such a distance governs its streaming quality such as delay.
Without considering such distances between peers in the physical network, the random peer selection mechanism used in many commercial P2P streaming systems are shown to be inefficient [3], [11]. Proximity-aware peer selection strategies [6], [4], [13] could remedy such inefficiency. In the proximity-aware P2P streaming systems, the peers are aware of such distance and select peers that are closer as their parents for downloading. The benefits of such proximity-aware mechanism in P2P streaming include (1) Reduced delay which is important for live streaming applications; (2) Reduced load on network by removing long-haul unicast connections, which also achieves ISP friendliness.
Towards the construction of high-quality P2P topologies, it is natural to ask the following questions for the proximity-aware streaming systems: (1) how server load and network bandwidth cost scale with the number of peers in the system; (2) how server load and network bandwidth cost scale with the delay tolerance of peers. Unfortunately, although the existing research have devised protocols to construct good proximity-aware P2P topologies, they fail to offer a comprehensive and analytical study on the characteristics that govern the scalability and performance of P2P streaming. On the other hand, though there exist analytical models for the P2P file sharing [18], [16], [22] and on-demand P2P streaming [21], [20], none of them could be applied to proximity-aware P2P streaming systems where the underlying physical network topology needs to be incorporated into the model.
In this paper, we seek analytical insights into the scalability of proximity-aware P2P streaming solutions. The challenge to incorporating topology concern into the P2P streaming analysis is evident from the complexity of Internet topology. To gain critical insights, we must construct an analytical model reasonably simple to derive closed-form results, meanwhile capturing the essential property of Internet topology. Towards this challenge, this paper proposes a novel H-sphere network model, which maps the network topology from the space of discrete graph to the continuous geometric domain. Our approach is motivated by the seminal study on power-law relation in Internet topology, which reveals the neighborhood size as a H-power function of hop distance [9].
Based on the H-sphere model, we perform in-depth analysis on a series of topology-aware peer selection methods and compare them with the random peer selection strategy. Our analytical investigation provides significate insights into the P2P streaming systems: First, of all peer selection methods studied, the server and network loads are independent of the peer population, but solely determined by the average outbound bandwidth of peers. Second, although random selection method can maximally save the server resource, it introduces the maximum load to the network.
The original contributions of this paper are two-fold. First, the novel H-sphere model enables in-depth analysis on topology-aware peer selection methods of different flavors. To the best of our knowledge, this is the first analytical study conducted in a topology-aware network setting. Second, we systematically investigate the proximity-aware P2P streaming strategies, by evaluating their performance via key scalability metrics, namely server load and network load. The analytical findings provide valuable guidelines for future P2P streaming system designs.
The remainder of this paper is organized as follows. We first present our H-sphere model for proximity-aware P2P streaming analysis in Section 2. Then we proceed to analyze the server load and network load in Sections 3 Server load analysis, 4 Network load analysis respectively. Finally, we validate our analytical model via a simulation-based study over Internet topologies in Section 5 and conclude the paper in Section 6.
Section snippets
H-Sphere model
To enable in-depth analysis on topology-aware peer selection methods, we first need to construct an analytical model reasonably simple to derive closed-form results, meanwhile capturing the essential property of Internet topology. We first characterize the distance between peers in the underlying physical network. Here the distance between nodes are measured by their hop count in the physical network, as it reveals many important performance metrics perceived by peers in the P2P system such as
Server load analysis
To derive the server load L, we need to know how much of the streaming workload is offset by peers. We do so by studying the amount of bandwidth received by each peer C from its supplying peers included in downloading region (the vertically shaded areas in Fig. 1).
Network load analysis
Now we turn to derive the network load M. M is defined as the summary of distances traveled by all data units within the network. The metric unit of M is the multiplication of bandwidth unit (such as Kbps) and distance, which is number of hops in topological networking terms, or geometric distance in the sphere model. Note that since the streaming bit rate is normalized to 1 in our analysis, M can be also regarded as the average delay, i.e., summary of peer-to-peer distance weighted by the
Simulation results
To validate our analytical observations obtained from the H-sphere model, we map them back to the real-world domain, and examine them via simulation over the topological network model, where the peer-to-peer distance is measured by the hop count of their shortest path.
Conclusion
In this paper, we present our analytical study on the impact of proximity-aware methodology to the scalability of P2P streaming. We propose a H-sphere model, which maps the network topology from the space of discrete graph to the continuous geometric domain, meanwhile capturing the the power-law property of Internet. Based on this model, we analyze a series of peer selection methods (random, variable-range, and fixed-range) by evaluating their performance via key scalability metrics, namely
References (23)
- Brite: Boston university representative internet topology generator. in: Available from:...
- Cooperative association for internet data analysis. in: Available from:...
- S. Ali, A. Mathur, H. Zhang. Measurement of commercial peer-to-peer live video streaming. in: Proceedings of ICST...
- S. Banerjee, B. Bhattacharjee, C. Kommareddy. Scalable application layer multicast. in: Proceedings of ACM SIGCOMM,...
- K.P.C.W.A.M.N. Carlsson. Analysis of bittorrent-like protocols for on-demand stored media streaming. in: Procceedings...
- Y. Chu, R. Rao, H. Zhang. A case for end system multicast. in: Proceedings of ACM SIGMETRICS,...
- et al.
Ostream: asynchronous streaming multicast in application-layer overlay networks selected areas in communications
IEEE J.
(2004) - Y. Cui, K. Nahrstedt. Layered peer-to-peer streaming. in: Proceedings of ACM NOSSDAV,...
- M. Faloutsos, P. Faloutsos, C. Faloutsos. On power-law relationships of the internet topology. in: Proceedings of ACM...
- et al.
Promise: peer-to-peer media streaming using collectcast
Cited by (6)
A survey on security in JXTA applications
2009, Journal of Systems and SoftwareModelNet-TE: An emulation tool for the study of P2P and traffic engineering interaction dynamics
2013, Peer-to-Peer Networking and ApplicationsMegaDrop: A cooperative video-on-demand system in a Peer-to-Peer environment
2011, Journal of Information Science and EngineeringA case for content distribution in peer-to-peer networks
2010, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)Probabilistic proximity-aware resource location in peer-to-peer networks using resource replication
2010, International Journal of Computers, Communications and ControlScalability and peer churning in IP-TV: An analytical insight
2009, GLOBECOM - IEEE Global Telecommunications Conference
- ☆
This work was supported by NSF award 0643488, Vanderbilt Discovery grant, and a gift from Microsoft Research. Views and conclusions of this paper are those of authors, which should not be interpreted as representing the official policies, either expressed or implied, of the funding agencies.