Section: New Results
Network measurement, modeling and understanding
The main objective of our work in this domain is a better monitoring of the Internet and a better control of its resources. In the monitoring part, we work on new measurement techniques that scale with the fast increase in Internet traffic and growth of its size. We proposed solutions for a fast and accurate identification of Internet traffic based on packet size statistics. We also studied the feasibility of inferring path metrics as the delay and the bandwidth from indirect measurements. In the network control part, we focus on new solutions that improve the quality of service to users by a better management of network resources and by a more efficient tuning of applications that take into account the constraints posed by the network. In this direction we propose efficient message drop and scheduling mechanisms for Disruption Tolerant Networks, as well as distributed topology-aware algorithms for the scheduling of communications among members of a wireless community interested in sharing data files among each other.
Next, is a sketch of our main contributions in this area.
Internet traffic classification by means of packet level statistics
One of the most important challenges for network administrators is the identification of applications behind the internet traffic. This identification serves for many purposes as in network security, traffic engineering and monitoring. The classical methods based on standard port numbers or deep packet inspection are unfortunately becoming less and less efficient because of encryption and the utilization of non standard ports. In this work we are looking for online iterative probabilistic methods that identify applications quickly and accurately by only using statistics on the sizes of packets. We want to associate a configurable confidence level to the port number carried in the transport header and to be able to consider a variable number of packets at the beginning of a flow. After preliminary verification on real traces, we observe that even in the case of no confidence in the port number, a very high accuracy can be obtained for well known applications after few packets were examined. A paper explaining our ideas and preliminary results is currently under submission.
An application-aware space for enhanced scalable services in overlay networks
We introduce the notion of an application aware space for enhanced scalable services in overlay networks. In this new space, the proximity of peers is determined according to a utility function that considers the network parameters (e.g., delay, bandwidth, and loss rate) impacting application performance. We motivate the need for this new notion by showing that the proximity in the delay space does not automatically lead to a proximity in another space (e.g., space of the bandwidth). For determining the proximity in this new space, network parameters must be estimated easily and scalably. Therefore, we use the matrix factorization approach for estimating the delay and loss parameters. Besides, we propose a scalable model that estimates the bandwidth among peers using the bandwidth of the indirect paths that join them via a set of landmarks. Our idea is that an indirect path shares the same tight link with the direct path with a probability that depends on the location of the corresponding landmark with respect to the direct path or any of the two peers subject to bandwidth inference. Our experimental results show that this new notion of proximity provides a much better quality than that obtained when using the delay proximity for large file transfer applications. The whole study is supported by real measurements carried out over Planetlab. The problem description and the results we obtained are summarized in  .
Understanding peer-to-peer dynamics
This axis focuses on the understanding and improvement of peer-to-peer content delivery. Indeed, we believe that the value of peer-to-peer comes from its ability to distribute contents to a large number of peers without any specific infrastructure, and within a delay that is logarithmic with the number of peers.
Following our previous results enabling a strong understanding of the BitTorrent core mechanisms, we have explored the practical issues that arise with the deployment of BitTorrent. In particular, in  , we have explored the impact of the piece size on the efficiency of BitTorrent. We have shown a non-trivial relationship between content size and piece size.
We have also worked, in the context of the Ph.D. thesis of Stevens Le Blond, on how to make BitTorrent ISP friendly  . One major issue with BitTorrent is that is does no take into account the underlying network topology. As a consequence some specific links are overloaded, and ISPs have to block BitTorrent traffic in order to decrease the load on those links. One solution to this problem is keep the BitTorrent traffic local to each ISP, leveraging on the ISPs network topology. This notion of locality has raised a huge interest recently. However, all proposed solutions consider only moderate locality. We have explored, running extensive very large scale experiments (with up to 10 000 peers on Grid5000), how BitTorrent behaves with high locality values (up to 99.998% ). We have shown that using such high locality values allows to reduce the cross ISP traffic up to two orders of magnitude without any significant impact of peers download completion time.
Finally, we continue to explore the impact of BitTorrent overlay structure on its performance. In particular, we have explored the impact of the strategy to build the overlay on BitTorrent efficiency. The considered strategies are the regular one using the tracker and a gossiping one called peer exchange. In addition to those strategies, we have introduced a variant called preemption. We perform our evaluation on large scale experiments on grid5000. This work in on-going and should lead to a technical report in 2009.