Section: New Results

Network measurement, modeling and understanding

Participants : Chadi Barakat, Walid Dabbous, Arnaud Legout, Mohammad Malli, Katia Obraczka, Clément Perrin, Karim Sbai, Thrasyvoulos Spyropoulos.

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. We also work on the utilization of measurements to infer the topology of the Internet and to localize any distributed resource. In the network control part, we focus on new solutions that improve the quality of service to users and that maximize the operators' revenues. Another objective is understanding the dynamics of the core mechanisms of peer-to-peer file sharing protocols. In particular, we focus on file transfer efficiency. We also want to investigate the existence and the potential impact of chaotic behaviors in the Internet. Understanding such behaviors is a scientific challenge, with the potential of a significant impact, especially concerning applications based, e.g., on chaos control, or resonances in chaotic systems.

Next, is a sketch of our main contributions in this area.

• Inferring Internet topology from application point of view

  We introduce in this work the notion of application-level proximity that serves for enhanced scalable services in overlay networks. This new proximity definition is a function of network parameters (e.g., delay and bandwidth) that decide on the application performance. We motivate the need for this new notion by showing that the network parameters are slightly correlated. Then, we consider two typical applications: a file transfer running over the TCP protocol, and an interactive audio service. For each application, we first propose a metric that models the application quality by considering the critical network parameters (e.g., delay, bandwidth, loss rate) affecting the application performance. Then, we evaluate the enhancement of the performance perceived by peers when they choose their neighbors based on our new proximity definition instead of the delay-based one determined using the proposed utility functions. Our major contribution is a model for inferring the bandwidth among peers in an easy and scalable manner. It consists of estimating the bandwidth among peers using the bandwidth of the indirect paths that join them via a set of well defined proxies or relays that we call landmark nodes. Our idea is that an indirect path shares the same tightest 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. We evaluate the impact of the location, number, and distribution of the landmarks on the bandwidth estimation accuracy. We obtain that the application-level proximity, which is determined using our bandwidth estimation model, provides much better quality than that obtained using the delay proximity for large file transfer applications. The whole study is supported by extensive measurements carried out over the worldwide experimental network Planetlab and is published in [3] , [29] , [30] .

• Reformulating the Monitor Placement Problem: Optimal NetworkWide Sampling

  Confronted with the generalization of monitoring in operational networks, researchers have proposed placement algorithms that can help ISPs deploy their monitoring infrastructure in a cost effective way, while maximizing the benefits of their infrastructure. However, a static placement of monitors cannot be optimal given the short-term and long-term variations in traffic due to rerouting events, anomalies and the normal network evolution. In addition, most ISPs already deploy router embedded monitoring functionalities. Despite some limitations (inherent to being part of a router), these monitoring tools give greater visibility on the network traffic but raise the question on how to configure a network-wide monitoring infrastructure that may contain hundreds of monitoring points. We reformulate the placement problem as follows. Given a network where all links can be monitored, which monitors should be activated and which sampling rate should be set on these monitors in order to achieve a given measurement task with high accuracy and low resource consumption? We provide a formulation of the problem, an optimal algorithm to solve it, and we study its performance on a real backbone network. This work is the result of a collaboration with Intel Research Cambridge, EPFL and Thomson. It is published in [14] .

• Modeling the AIADD Paradigm in Networks with Variable Delays

  Modeling TCP is fundamental for understanding Internet behavior. The reason is that TCP is responsible for carrying a huge quota of the Internet traffic. During last decade many analytical models have attempted to capture dynamics and steady-state behavior of standard TCP congestion control algorithms. In particular, models proposed in literature have been mainly focused on finding relationships among the throughput achieved by a TCP flow, the segment loss probability, and the round trip time (RTT) of the connection, which the flow goes through. Recently, Westwood+ TCP algorithm has been proposed to improve the performance of classic New Reno TCP, especially over paths characterized by high bandwidth-delay products. We we developed an analytic model for the throughput achieved by Westwood+ TCP congestion control algorithm when in the presence of paths with time-varying RTT. The proposed model has been validated by using the ns-2 simulator and Internet-like scenarios. Validation results have shown that this model provides relative prediction errors smaller than 10%. It has been shown that a similar accuracy is achieved by analogous models proposed for New Reno TCP. Moreover, it has been proved that it is necessary to consider delay variability in modeling Westwood+ TCP; otherwise, if only the average RTT is considered, performance could be underestimated. All results can be found in [13] . This work was done with the Maestro group at INRIA Sophia Antipolis and with Politecnico di Bari in Italy.

• TICP: Transport Information Collection Protocol

  We worked on improving and validating TICP [7] , our TCP-friendly reliable transport protocol to collect information from a large number of sources spread over the Internet. A collector machine sends probes to information sources, which respond by sending back report packets containing their information. TICP adapts the rate of probes in a way to avoid implosion at the collector and congestion in the network. To ensure smooth variation of the congestion control parameters and to probe sources behind the same bottleneck at the same time, we add to TICP a mechanism that clusters information sources. This mechanism is based upon the Global Network Positionning (GNP) Internet coordinate system. By running simulations in ns-2 over realistic network topologies, we prove that TICP with clustering of information sources has shorter collect session duration and causes less packet losses than the initial version that probes sources independently of their location.

• Understanding peer-to-peer dynamics

  We started a collaboration with Pietro Michiardi and Guillaume Urvoy-Keller from the Institut Eurecom. We instrumented a BitTorrent client and performed large scale experiments to understand the dynamics of the core BitTorrent mechanisms. Such an experimental study was never performed before. Indeed, the previous studies of BitTorrent were based either on simulations or modeling; and these studies presented important restrictions. We evaluated BitTorrent's two core mechanisms: its piece selection mechanism called rarest first, and its peer selection algorithm called choke algorithm [26] . We show that the rarest first algorithm guarantees a diversity of the pieces among peers close to the ideal one. In particular, on our experiments, a replacement of the rarest first algorithm with a source or network coding solution cannot be justified. We also show that the choke algorithm in its latest version fosters reciprocation and is robust to free riders. In particular, the choke algorithm is fair and its replacement with a bit level tit-for-tat solution is not appropriate.

  Then we started a collaboration with Nikitas Liogkas, Eddie Kohler, and Lixia Zhang from UCLA, USA. Focusing on the properties of the choke algorithm [36] , we show that it enables clustering of similar-bandwidth peers, ensures effective sharing incentives by rewarding peers who contribute with high download rates, and achieves high upload utilization for the majority of the download duration. We also examine the properties of the new choke algorithm in seed state and the impact of initial seed capacity on the overall BitTorrent system performance. In particular, we show that an underprovisioned initial seed does not enable clustering of peers and does not guarantee effective sharing incentives. However, we show that even in such a case, the choke algorithm guarantees an efficient utilization of the available resources by enforcing fast peers to help other peers with their download. Based on their observations, we offer guidelines for content providers regarding seed provisioning, and discuss a tracker protocol extension that addresses an identified limitation of the protocol. Those results are available in a technical report [36] that is under submission.

• Chaotic behavior in computer networks

  Chaos is a prominent feature of complex systems and dynamical systems theory provides methods for analyzing this kind of behavior. Computer networks are complex systems, but it is not yet known whether Internet protocols exhibit chaotic behaviors though some preliminary investigations suggest it. Understanding the meaning and effect of such behaviors is a scientific challenge, with the potential of a significant impact, especially concerning applications based, e.g., on chaos control, or resonances in chaotic systems. For this, on the one hand, one needs to design models describing properly the dynamic evolution of a computer network using an Internet protocol, and to make the mathematical analysis of these models. On the other hand, one must perform careful investigations on real protocol traces, to analyze them with the tools developed in chaos theory, and to compare them to the predictions of the models. We have recently started a collaboration with researchers from the Institut Non-Linéaire de Nice (INLN) on this topic. We received a grant (COLOR CAOREDO) for one year to start this research topic, and an internship was made on that subject (Clément Perrin, 2006).

• Disruption Tolerant Networking

  We start an activity on problems related to efficient routing in Delay Tolerant Networks (DTN). DTNs are networks where a number of traditional assumptions break, and novel communication techniques need to be applied. Two of these techniques that have found considerable success are that of "mobility-assisted routing" and "controlled message replication". We have already identified a family of protocols, called Spray routing, who successfully combine these techniques and achieve close-to-optimal performance in idealized scenarios, where for example all nodes are homogeneous or all nodes are co-operating in forwarding traffic.

  We're currently investigating the performance of these protocols under non-ideal conditions including: losses of message replicas (e.g. queue drops, non-cooperating nodes, etc.), heterogeneous environments, correlated mobility in both time and space. We focus on both evaluating/modeling the performance degradation of these protocols compared with the ideal situation, as well as designing new mechanisms (including history-based algorithms, learning, adaptability) that can overcome these problems and deliver superior performance in diverse conditions.

  Also related to networking in environments subject to episodic connectivity, we have been working on the following activities: (1) Developing a taxonomy for existing mechanisms and protocols, (2) exploring different heuristics for opportunistic message forwarding, and (3) investigating the effects of traffic differentiation on the performance of different routing mechanisms.