Shao Liu (Julian ) -> Research -> RTT Variation Modeling

HOME RESEARCH BACKGROUND PERSONAL LINKS

 

Modeling on RTT Variations in Fluid Model

One power for the fluid model is that we can use control theory to analyze the stability property of a congestion control system. One difficulty of analyzing stability is that the system is a delayed system, i.e., the plant gets the delayed control signal. Considering this delay, it is very challenging to analyze the global stability, and one standard technique is using generalized Nyquist method to analyze the local stability. The main idea is that we find the relationship of the perturbations of the state variable (send rate) and the control signal (congestion signal) around their respective equilibrium values, and the non-linear system is linearized around the operating point.

This method also faces one problem: the linearization of RTT. Sometimes RTT are treated as a variable, and sometimes it is treated as a constant, and we have shown that we need to be very cautious in this RTT modeling: different RTT modeling may lead to completely different results on stability property, and the variance of RTT may influence the stability in different directions:

  1. For RED, the variable RTT nature makes it easier to be stabilized, and if RTT is constant, the stability region is smaller than if RTT is a variable
     
  2. For delay based algorithms, like TCP-Vegas, Stabilized Vegas, the variable nature of RTT make these algorithms harder to be stabilized. When RTT is treated as a constant, Stabilized Vegas can be stable, but if we consider the RTT variation, we see that Stabilized Vegas, together with Vegas, are unstable when the number of bottleneck links exceed 4.

Papers on this project:

[C2] S. Liu, T. Başar and R. Srikant. “Pitfalls in the Fluid Modeling of RTT Variations in Window-based Congestion Control”Proc. of IEEE Infocom, Miami, FL, March 2005. 

The slides is available here.

[J3] S. Liu, T. Başar and R. Srikant. “Modeling Delay in Window-based Congestion Control”, to be submitted to IEEE Transactions on Automatic Control.

 

Related Work

Towsley, etc's analysis on the instability of RED with TCP

TCP-Vegas

Stabilized Vegas

 

Shao Liu