Section: New Results
Congestion control and IP traffic characterization
Participants : Sara Alouf, Konstantin Avrachenkov, Alberto Blanc, Abdulhalim Dandoush, Alain Jean-Marie, Philippe Nain, Giovanni Neglia.
Evaluation of new TCP versions
Participants : Konstantin Avrachenkov, Alberto Blanc, Giovanni Neglia.
Compound TCP is one of the many new versions of TCP for high speed networks. In [64] , [115] A. Blanc, K. Avrachenkov and D. Collange (Orange Labs ) use a fluid model to analyze the behavior of an isolated Compound TCP connection.
In [63] , [116] , K. Avrachenkov, A. Blanc and G. Neglia, together with D. Collange (Orange Labs ) study the performance of Compound TCP under random losses. Markovian and deterministic models are used to derive the steady state distribution of the window along with synthetic performance metrics like average throughput and coefficient of variation of the window for an immediate comparison with TCP Reno.
In [62] A. Blanc, K. Avrachenkov and D. Collange (Orange Labs ) use a Markovian model to compare the performance of different TCP versions (Compound TCP, Cubic, HighSpeed, and Reno) under Bernoulli losses, computing the average window size (response function), the Coefficient of Variation (CoV) of the window and the average throughput.
Estimating the round-trip time of long-lived TCP sessions
Participants : Sara Alouf, Konstantin Avrachenkov, Alberto Blanc, Philippe Nain.
The Round-Trip Time (RTT) of a TCP connection represents an important characteristic whose knowledge is useful when controlling a long-lived flow at a router. In [132] , K. Avrachenkov, S. Alouf, and P. Nain, in collaboration with D. Carra (University of Verona , Italy) and G. Post (Alcatel-Lucent Bell Labs ), propose a passive, online RTT estimation methodology based on the traffic observed in one direction. The method uses spectral analysis along with a pattern-matching technique for the extraction of the fundamental frequency. Since the proposed solution estimates in real-time the RTT using one-way traffic, it represents a candidate for a possible implementation in routers.
Together with A. Blanc, the authors have validated the methodology through measurements in a controlled testbed and on the Internet. The results can be found in [117] .
This research is carried out within ADR “Semantic Networking” (see Section 7.1.1 ).
Flow-aware traffic management
Participants : Sara Alouf, Konstantin Avrachenkov, Alberto Blanc.
The congestion control mechanism of TCP, while simple and scalable, has several well-known limitations: 1) often different flows experience synchronized losses leading to lower link utilization, and 2) when flows with different Round-Trip Times (RTT) share the same bottleneck link, flows with a smaller RTT receive a larger share of the capacity.
In [103] , A. Blanc, K. Avrachenkov, and S. Alouf, in collaboration with G. Post (Alcatel-Lucent Bell Labs ), propose a new flow-aware traffic management mechanism that aims at addressing the two aforementioned limitations, while being self-configuring and supporting different fairness criteria. The core idea of the proposed mechanism can be described as a two step process: 1) decide a target rate for each flow; 2) control each TCP flow in order to minimize the oscillations around the chosen target rate.
This research is carried out within ADR “Semantic Networking” (see Section 7.1.1 ).
Remote active queue management
Participant : Eitan Altman.
In [75] E. Altman, in collaboration with M. Ibrahim and P. Vicat-Blanc Primet (Inria project-team Reso ), G. Carofiglio and G. Post (Alcatel-Lucent Bell Labs ), has proposed an Active Queue Management (AQM) scheme that is able to detect congestion in other nodes. By reacting to that congestion in its own node, it is able to reduce the congestion in the other nodes. The detection of congestion is obtained through the identification of a change in the rate of increase of the congestion window of TCP connections. The authors study through simulations the accuracy of this way to detect congestion.
Flow-level simulation of parallel downloads
Participants : Abdulhalim Dandoush, Alain Jean-Marie.
Parallelism in the download process of large files is an efficient mechanism for distributed systems. In such systems, some peers (clients) exploit the power of parallelism to download blocks of data stored in a distributed way over some other peers (servers). Parallel downloading with capacity constraints on both the client downloads and server uploads has not been well analyzed. In particular, a basic problem is to predict the instantaneous shares of the bandwidths of each client/server devoted to each data transfer flow. A. Dandoush and A. Jean-Marie have proposed and analyzed a simple algorithm that works at the flow-level and uses the concept of “Water-Filling” (or min-max fairness). The response times of parallel downloading have been analyzed (both distributions and averages) using the algorithm by flow level simulations. The results have been compared to those of packet-level simulations after implementing the same process in Network Simulator ns-2.
