Parallel TCP Sockets

Overview

We found a simple throughput formula for N identical TCP connections over a bottleneck of capacity c in absence of TCP window synchronization:

x(N) / c = 1 – 1 / (1+3N)

This simple formula appears to be previously unknown. Its proof follows under mild “stability conditions”. For N = 1, it boils down to well known TCP throughput formula under periodic loss events. The formula may be deemed fairly general in that it accommodates variety polices to select which connection is signalled to undergo window reduction. These loss policies may be emergent from dynamics in networks or may be enforced by intelligent dropping elements. The formula tells us that TCP throughput inefficiency due to TCP window adaptation in congestion avoidance is effectively already eliminated with a few parallel TCP connections, for example, it predicts that 3 connections would achieve 90% utilization, while 6 connections achieve almost 95%. This would provide incentives to throughput-greedy users not to open too many parallel TCP connections as already a few would be enough. Other factor that results in throughput degradation is TCP window synchronization, which may provide incentive to users to open more sockets in order to improve their throughput. But this is another factor.

The result may be of interest to concurrent community of users that use parallel TCP sockets for bulk data transfers, with standard protocols such as GridFTP and bbftp and several shareware "download manager" software. This analysis result is in accordance with empirical results, e.g. Hacker et al and soon we will post our own experimental evidence. The result also suggests that MultTCP, which was designed to achieve mean throughput of a given number of virtual TCP flows, yields a very good approximation.

The formula tells more: the aggregate throughput is insensitive on particular choice of loss policy, under mild stability conditions.

Publications

• Parallel TCP Sockets: Simple Model, Throughput and Validation, E. Altman, D. Barman, B. Tuffin, and M. Vojnovic, (MSR-2005-130), Sep 2005, (shorter version presented at IEEE Infocom 2006).

Slides

• Infocom 2006 talk slide deck.

People

Eitan Altman, Dhiman Barman, Bruno Tuffin, Milan Vojnovic

Related Work

• TCP Modelling by Baccelli and Hong (see the references therein).
• "The end-to-end performance effects of parallel TCP sockets on a lossy wide-area network", T. Hacker, B. Athey, B. Noble, IEEE-CS/ACM International Parallel and Distributed Processing Symposium, Ft. Lauderdale, FL, April, 2002.