M.Sc Thesis

M.Sc StudentTavor Eran
SubjectTCP Congestion Control Mechanism Modeling and Analysis
DepartmentDepartment of Electrical and Computer Engineering
Supervisor PROFESSOR EMERITUS Moshe Sidi


TCP is one of the most popular protocols in the Internet. Its performance has been widely investigated over the past few years. This work presents a detailed performance model for TCP’s congestion control mechanism which captures not only the Congestion Avoidance (CA) phase of the protocol but also its Loss Recovery (LR) phase. Furthermore, this is done not only in presence of Bernoulli distributed packet loss model but also in presence of Gilbert bursty errors which introduce correlation between the lost packets.

Early studies of TCP examined the AIMD (Additive Increase Multiplicative Decrease) Nature of the CA (Congestion Avoidance) phase of TCP. These early studies yielded the famous formula  where TR is the transmission rate, MTU is the Maximum Transfer Unit, RTT is the Round Trip Time of the connection and Ploss is the probability that a packet will be lost. Important assumptions are that the connection has a large bandwidth delay product and that the losses are Bernoulli distributed. Later on the calculations were extended to include the influence of timeouts (TO).

We present here our model which is a discrete time and finite Markov chain. This model captures not only Congestion Avoidance (CA) and timeouts, but also correlation between the lost packets. Our analysis shows that the throughput of a TCP connection degrades when correlation is introduced (while the average loss rate is maintained).

In order to validate our model we used the NS2 network simulator. We also present a TCP congestion control survey in appendix A. In this survey we cover the various methods for congestion control suggested for TCP over the years.