Team Gang

Overall Objectives
Scientific Foundations
Application Domains
New Results
Contracts and Grants with Industry
Other Grants and Activities

Section: New Results


Video-on-Demand over Peer-to-Peer networks

Participants : Yacine Boufkhad, Fabien Mathieu [ Orange Labs, Issy Les Moulineaux, France ] , Fabien de Montgolfier, Diego Perino, Laurent Viennot.

We consider the fully distributed Video-on-Demand problem, where n nodes called boxes store a large set of videos and collaborate to serve simultaneously n videos or less between them. It is said to be scalable when $ \upper_omega$(n) videos can be distributively stored under the condition that any sequence of demands for these videos can always be satisfied. Our main result consists in establishing a threshold on the average upload bandwidth of a box, above which the system becomes scalable. We are thus interested in the normalized upload capacity Im7 ${u=\mfrac {upload~bandwidth}{video~bitrate}}$ of a box. The number m of distinct videos stored in the system is called its catalog size.

In [7] , we show an upload capacity threshold of 1 for scalability in a homogeneous system, where all boxes have the same upload capacity. More precisely, a system with u<1 has constant catalog size m = O(1) (every box must store some data of every video). On the other hand, for u>1 , an homogeneous system where all boxes have same upload capacity at least u admits a static allocation of m = $ \upper_omega$(n) videos into the boxes such that any adversarial sequence of video demands can be satisfied. Moreover, such an allocation can be obtained randomly with high probability. This result is generalized to a system of boxes that have heterogeneous upload capacities under some balancing conditions.

In [8] , by means of extensive simulations we analyze the impact of: i) the video allocation technique used for distributed storage ii) the use of cache to allow nodes to re-distribute the video they are downloading iii) the use of static/dynamic algorithms for video distribution.

Based on these results, we provide some guidelines for setting the system parameters: the use of cache strongly improves system performance; popularity based allocation techniques can be sensitive and bring little improvement; dynamic distribution algorithms are needed only in extreme scenarios while static ones are commonly sufficient.

Unstructured P2P live streaming

Participants : Nidhi Hegde [ Orange Labs, Issy Les Moulineaux, France ] , Fabien Mathieu [ Orange Labs, Issy Les Moulineaux, France ] , Diego Perino [ Orange Labs, Issy Les Moulineaux, France ] .

In unstructured P2P live streaming systems the stream is not forwarded as a continuous flow of data but is divided in a series of pieces (chunks), that are injected in the system by a source and exchanged among peers in order to retrieve the complete sequence and play out the stream. Data exchange is therefore driven by chunk exchange algorithms run locally by nodes, which can be described by their chunk/peer selection policies. It turns out that the most popular commercial peer-to-peer systems for live streaming like CoolStreaming, PPLive, SopCast, are based on such an unstructured approach.

The performance trade-offs of chunk exchange algorithms have been deeply analyzed for homogeneous systems, where all peers have the same upload capacity. In [27] we analyze chunk diffusion algorithms designed for heterogeneous environments, where peers have different upload capacities. We focus on the peer selection process and propose a generic model that encompasses a large class of algorithms. We derive recursive formulas to describe the chunk diffusion function of a generic latest blind chunk / resource aware peer selection scheme. By means of simulations, we analyze the resource awareness-agnostism trade-offs on the peer selection process and the impact of the source distribution policy in non-homogeneous networks. We highlight that the early diffusion of a given chunk is crucial for its overall diffusion performance, and a fairness trade-off arises between the performance of heterogeneous peers, as a function of the level of awareness. Moreover, we show the critical role the source selection policy plays on chunk diffusion performance.

A good diffusion scheme is indeed essential for the performance of an unstructured P2P live streaming system. For a given scheme however, an optimization at a detailed level is also important. This involves the fine tuning of dissemination parameters, such as chunk size, receiver buffer size, number of peers to probe, etc. In [26] , [23] we investigate optimal sizing of chunks and probe sets, i.e. the number peers a given node probes before transmitting chunks. The analysis is performed by means of an event-based simulator. We show that the chunk size significantly impacts performance and that it should fall within a given range which is mostly determined by the median RTT of the network and the stream rate. We also show that the size of the probe set affects performance of diffusion schemes, and, in particular, a probe set larger than the actual number of concurrent connections may improve miss ratio/delay performance by modifying the suitable chunk size ranges.


Logo Inria