Section: New Results

High Speed Network's traffic metrology and statistical analysis

Impact of the Correlation between Flow Rates and Durations on the Large-Scale Properties of Aggregate Network Traffic

Keywords : heavy-tailed flow size distributions, long range dependence, aggregated network traffic, correlations.

Participants : Patrick Loiseau, Paulo Gonçalves, Pascale Vicat-Blanc Primet.

Since the discovery of long-range dependence in network traffic in 1993, many models have appeared to reproduce this property, based on heavy-tailed distributions of some flow-scale properties of the traffic. However, none of these models consider the correlation existing between flow rates and flow durations. In this work, we extend previously proposed models to include this correlation. Based on a planar Poisson process setting, which describes the flow-scale traffic structure, we analytically compute the auto-covariance function of the aggregate traffic's bandwidth and show that it exhibits long-range dependence with a different Hurst parameter. In uncorrelated case, the model that we propose is consistent with existing models, and predict the same Hurst parameter. We also prove that pseudo long-range dependence with a different index can arise from highly variable flow rates. The pertinence of our model choices is validated on real web traffic traces.

Maximum likelihood estimate of heavy-tail exponents from sampled data

Keywords : heavy-tail distributions, maximum likelihood estimation, flow size.

Participants : Patrick Loiseau, Paulo Gonçalves, Pascale Vicat-Blanc Primet.

This work, published in the proceedings of ACM Sigmetrics 2009 [36] , is a joint collaboration with the MISTIS team project.

In the context of network traffic analysis, we address the problem of estimating the tail index of flow (or more generally of any group) size distribution from the observation of a sampled population of packets (individuals). We give an exhaustive bibliography of the existing methods and show the relations between them. The main contribution of this work is then to propose a new method to estimate the tail index from sampled data, based on the resolution of the maximum likelihood problem. To assess the performance of our method, we present a full performance evaluation based on numerical simulations, and also on a real traffic trace corresponding to internet traffic recently acquired.

A new model revealing unexplored scale invariant properties of TCP throughput

Keywords : long-lived TCP flow, multifractal analysis, large deviation principle, markov model.

Participants : Patrick Loiseau, Paulo Gonçalves, Pascale Vicat-Blanc Primet.

This is a joint work with J. Barral (Prof. Univ. Paris 13).

Classical scaling laws in network traffic are commonly accepted as valuable indicators of the system's state. However, they are usually related to the dynamic of the connections, rather than to the TCP control mechanism itself. In this work we identified new scale-invariance properties of the throughputs of long-lived TCP RENO connections, and we proposed an adequate model able to reproduce these scaling laws. Our model relies on simple Markov chains for which we can theoretically prove, and experimentally verify, that they inherently possess the sought properties. We derived the corresponding large deviation spectra, which reveal reliable and sensitive fingerprints of the performance and fairness of competing TCP flows. Under controlled experimental conditions, we then demonstrated the flexibility and the versatility of this original approach on real TCP traces. In particular, we showed that the specificities of experimental conditions, such as synchronization or cross-traffic nature, can easily be taken into account and embedded in the model. We also presented experimental evidence that different TCP variants also exhibit scale-invariant properties of the same kind.

Traffic classification techniques supporting semantic networks

Keywords : semantic networking, supervised classification, elephants and mice.

Participants : Olivier Grémillet, Paulo Gonçalves, Pascale Vicat-Blanc Primet.

This work is part of our activity within Common Lab between INRIA and Alcatel-Lucent Bell Labs; it has been carried out in close collaboration with A. Dupas (Alcatel-Lucent).

The Semantic Networking concept has been introduced to solve the QoS, scalability and complexity challenges for the Future of Internet. Based on traffic awareness and flow entity, it contributes to an adaptive management of the network. The first important features are the better knowledge of the transported traffic and the processing time of the classification compatible with real-time operation. In this work, we present interesting techniques of classification for semantic networks. The detection of the biggest flows is first studied with Sample and Hold and multi-stage filter methods with successful classification probability. We hence analyze the impact of flow parameters on the application identification performance and classify them according to their accuracy. We finally discuss a potential hardware implementation architecture to validate the concept of semantic networking.

Multifractal analysis for in partum fetal-ECG diagnosis

Keywords : Multifractal analysis, wavelets, heart rate variability, detection, supervised classification.

Participant : Paulo Gonçalves.

Albeit grounded on different physical origins, it is not rare that distinct real-world problems share common mechanisms and/or formulations. This similitude naturally fosters the development of unified frameworks which can then match a wide range of applications. Moreover, as statistical signal processing frequently stands at the junction of several scientific domains, it is not surprising that statistical studies go beyond the scope of the application areas they were initially addressing. That is why the RESO team was led to participate to inter-disciplinary collaborations that do not straightforwardly relate to the main themes of the project activities, but which can capitalize with them.

This is a joint work with the Sisyphe team of the ENS Lyon Physics Lab and with the obstetric group of Hôpital Femme Mère Enfant of Hospices civils de Lyon .

In partum fetal suffering surveillance is a key task to prevent fetal and neonatal mortality due to asphyxia. This is partially conducted by monitoring and analyzing fetal Electrocardiogram recorded during the delivery phase: A strong variability measured on the corresponding heart beat time series indicates a normal process. Though satisfactory in practice, the currently used analysis/decision criteria lead to a high number of false positives. Multifractal analysis can be envisaged as a new tool to revisit time series variability analysis. Applied to data collected at Hospices Civils de Lyon, France, wavelet Leader based multifractal analysis is shown here to achieve a significant discrimination between the True Negative, True positive and False Positive classes of patients. This hence open promising tracks to decrease the number of False Positives achieved.