TCP Stacks on production links, Fast TCP and Reno TCP single streams varying the TCP window size  Main Page of TCP Stacks on production links | Summary of single stream tests | High Speed TCP single streams versus a Reno TCP single stream | Scalable TCP single streams versus a Reno TCP single stream |  Scalable TCP single stream versus Reno TCP multiple streams | Fast TCP single stream versus a Reno TCP single stream | Fast TCP single stream versus Reno TCP multiple streams | Bulk throughput measurements Fabrizio Coccetti and Les Cottrell. Created 7 April '03, last update 7 April '03

IEPM Home Page  IEPM Site Map  Network Monitoring  PingER  IEPM-BW Home IEPM Papers and Presentations

Overview | Reno TCP | Fast TCP | Fast TCP vs. Reno TCP

Overview

In this page we run tests using two single streams from two different machines each using a different TCP stack to the same destination machine.
All measures have a standard MTU (1500B) and txqueuelen (100packets), the TCP window has values: 8MB, 16MB and 32MB.
Tests run on Linux machines, the window size we indicate is the requested window size, i.e. it is the one you use with the Iperf command via the switch -w.

All these tests have been performed from SLAC to CERN, on April 4th, 2003 from 6 pm to 10 pm.

Measures of competing streams An Iperf single stream is started from S1 to the Remote Host. At the same time, another Iperf single stream is started from S2 the Remote Host. We display the values of the throughput of the two streams every 5 seconds. Measure of non-competing streams We measure the performance of Fast TCP starting an Iperf single stream from S1 to the Remote Host.  We measure the performance of Reno TCP starting an Iperf single stream from S2, after the test from S1 has finished.

Reno TCP, from SLAC to CERN, window size of 8MB, 16MB, 32MB

Tests show that a Reno TCP single stream from SLAC to CERN reaches the best performance using a TCP window of 16 MB.  All test have been performed in the evening on April 4th, 2003.

	Reno TCP Throughput Average = 164.6 Mbps Std Dev = 37.4 Mbps RTT RTT Average = 207 ms Std Dev = 2 ms
	Reno TCP Throughput Average = 221.5 Mbps Std Dev = 64.3 Mbps RTT RTT Average = 207 ms Std Dev = 0 ms
	Reno TCP Throughput Average = 190.7 Mbps Std Dev = 117.6 Mbps RTT RTT Average = 209 ms Std Dev = 14 ms

Fast TCP, from SLAC to CERN, window size of 8MB, 16MB, 32MB

Tests show that a Fast TCP single stream from SLAC to CERN reaches the best performance using a TCP window of 16 MB.  The average throughput (for a 20 min measure) is more than 20 % greater than the throughput obtained with a TCP window of  8MB or 16MB. All test have been performed in the evening on April 4th, 2003.

	Fast TCP Throughput Average = 239.7 Mbps Std Dev = 97.9 Mbps RTT RTT Average = 208 ms Std Dev = 4 ms
	Fast TCP Throughput Average = 294.4 Mbps Std Dev = 66.3 Mbps RTT RTT Average = 208 ms Std Dev = 7 ms
	Fast TCP Throughput Average = 237.6 Mbps Std Dev = 95.8 Mbps RTT RTT Average = 208 ms Std Dev = 3 ms

Fast TCP vs. Reno TCP (competing each other), from SLAC to CERN, window size of 8MB, 16MB, 32MB

Tests show a Reno TCP single stream competing with a Fast TCP single stream from SLAC to CERN.
The aggregate throughput reaches the best performance using a TCP window of 16 MB.  All test have been performed in the evening on April 4th, 2003.

	Fast TCP Throughput Average = 226.7 Mbps Std Dev = 98.4 Mbps Reno TCP Throughput Average = 21.6 Mbps Std Dev = 31.0 Mbps RTT RTT Average = 209 ms Std Dev = 14 ms
	Fast TCP Throughput Average = 285.5 Mbps Std Dev = 77.1 Mbps Reno TCP Throughput Average = 44.7 Mbps Std Dev = 30.0 Mbps RTT RTT Average = 209 ms Std Dev = 11 ms
	Fast TCP Throughput Average = 237.6 Mbps Std Dev = 95.8 Mbps Reno TCP Throughput Average = 58.4 Mbps Std Dev = 63.3 Mbps RTT RTT Average = 208 ms Std Dev = 6 ms