Traffic Awareness (Traffic)
Leader: Paulo Gonçalves (RESO)
Metrology based decision should become an inherent component of future Internet. Having in mind, admission control, bandwidth sharing, differentiated flow treatment, congestion avoidance, to cite but a few possible objectives, we propose to address the challenging task of traffic analysis and characterization from two directions.
Semantic Networking Our aim is to identify relevant features of traffic data able to discriminate between the underlying classes of applications. The ultimate goal is then to condition differentiated treatments to different flow categories, having in mind to optimize resources allocation and sharing. To this end, our work encompasses several facets : sampling theory, statistical characterization and modeling of discrete time series (i.e. point processes), parameter estimation, learning theory for classification purpose. All these tasks must comply with real time constraints, to guarantee that traffic awareness can be effectively implemented in the core routers of very high speed networks.
Performance evaluation An important advantage of the metrology testbed we strive at promoting, is its ability to generate, under realistic network situations, real lived traffics whose origin and transfer conditions are fully controlled. In particular, we can monitor the buffer occupancy at a bottleneck point and then relate the congestion characteristics (global and per flow loss rates, packets? drop distribution, transmission delay estimation, etc) to statistical ? or geometrical ? properties of the input traffic. More precisely, we are concerned with scaling properties of aggregated and flow based throughputs (adapting to effectively impacting time scales what has been done with long range dependence), and on their dependence on the used protocols. This study led us to consider Markov models of TCP traffics and to devise a new multifractal analysis approach to account for the particular time varying instantaneous throughput due to retroactive control mecha- nisms. We then expect to find some mapping between this large deviation principle and characteristics of TCP variants, which would explain different measured performances.