Modeling TCP Throughput with Random Packet Drops

Abstract

The present report deals with the modeling of the long-term throughput, a.k.a., send rate, of the Transmission Control Protocol (TCP) under the following assumptions. (i) We consider a single 'infinite source' using a network path from sender to receiver. (ii) Each TCP packet is randomly dropped with probability p; independently of previous drops or any other event/parameter. (iii) The - never changing - receiver window limits the amount of outstanding data. (iv) The receiver acknowledges every packet. (v) The TCP modeled here conforms to the publicly available standards (RFCs) as concerns congestion control. We validate and determine the limits of the different models proposed here using packet-level simulations. The contributions of the present work are the following: (a) We determine three regimes, and their conditions of applicability, depending on p: Linear law regime, square root law regime, and timeout regime. (b) As concerns the relationship between the linear and square root regimes, we give additional insights relatively to previously published work. (c) We give the exact equations governing the TCP send rate in any regime. (d) From the exact equation and under the further condition that the path is not saturated, we give and discuss approximations for the send rate of the NewReno variant of TCP. A by-product of these calculations is the distribution of the sender window, independently of any timing or saturation consideration. (e) These approximations give results that are accurate to a few percent when compared to simulation results. Detailed comparison and sources of errors between theory and simulations are also discussed.