An asymptotic approximation for TCP CUBIC

In this paper, we derive an expression for computing the average window size of a single TCP CUBIC connection under random losses. For this we consider a throughput expression for TCP CUBIC computed earlier under deterministic periodic packet losses. We validate this expression theoretically. We then use insights from the deterministic loss-based model to scale appropriately a sequence of Markov chains with random losses indexed by the probability of loss p. We show that this sequence converges to a limiting Markov chain as p tends to 0. The stationary distribution of the limiting Markov chain is then used to derive the average window size for small packet error rates. We then use a simple approximation to extend our current results with negligible queuing to a setup with multiple connections and non-negligible queuing. We validate our model and approximations via simulations.