TCP-Illinois: A loss and delay-based congestion control algorithm for high-speed networks

We introduce a new congestion control algorithm, called TCP-Illinois, which has many desirable properties for implementation in (very) high-speed networks. TCP-Illinois is a sender side protocol, which modifies the AIMD algorithm of the standard TCP (Reno, NewReno or SACK) by adjusting the increment/decrement amounts based on delay information. By using both loss and delay as congestion signals, TCP-Illinois achieves a better throughput than the standard TCP for high-speed networks. To study its fairness and stability properties, we extend recently developed stochastic matrix models of TCP to accommodate window size backoff probabilities that are proportional to arrival rates when the network is congested. Using this model, TCP-Illinois is shown to allocate the network resource fairly as in the standard TCP. In addition, TCP-Illinois is shown to be compatible with the standard TCP when implemented in today's networks, and is shown to provide the right incentive for transition to the new protocol. We finally perform ns-2 simulations to validate its properties and demonstrate its performance.