Tags:
create new tag
view all tags

Throughput Tests with ASGC

These throughput tests were done with Jason Shih, Chia Ming Kuo and Aries Hung.

The doIperf.sh was run from a Linux 2.6 kernel system

Linux ce03-lcg.projects.cscs.ch 2.6.9-42.0.10.EL.cernsmp #1 SMP 
             Thu Mar 1 15:11:46 CET 2007 i686 athlon i386 GNU/Linux

The throughput looks much better than the one from the test of April 30th done with the Linux 2.4 kernel system. But a test measurement on the same Linux 2.4 system done immediately after the Linux 2.6 measurements showed the same improvements (on the same page). The variation of the TCP congestion algorithm also has no great effect on the upload to ASGC, as can be seen from the graphs below.

Basic sysctl settings

The default tcp window size settings are much too small, so changed them to

sysctl -w net.ipv4.tcp_rmem="131072     1048576      2097152"
sysctl -w net.ipv4.tcp_wmem="131072     1048576      2097152"
sysctl -w net.ipv4.tcp_mem="131072     1048576      2097152"
sysctl -w net.core.rmem_default=1048576
sysctl -w net.core.wmem_default=1048576
sysctl -w net.core.rmem_max=2097152
sysctl -w net.core.wmem_max=2097152

Route, RTT and packet loss

HOST                                    LOSS  RCVD SENT    BEST     AVG   WORST
148.187.33.2                              0%   100  100    0.37    0.51    2.13
148.187.32.4                              1%    99  100    0.10    0.21    0.55
swima2.cscs.ch                            0%   100  100    0.34    2.13  162.41
swiEL2-10GE-1-4.switch.ch                 0%   100  100    3.09    3.33    4.05
swiCE3-10GE-1-3.switch.ch                 0%   100  100    3.84    9.98  185.01
swiCE2-10GE-1-4.switch.ch                 0%   100  100    3.93    4.80   58.08
switch.rt1.gen.ch.geant2.net              0%   100  100    3.97    4.81   37.65
so-7-2-0.rt1.fra.de.geant2.net            1%    99  100   12.11   12.63   30.43
abilene-wash-gw.rt1.fra.de.geant2.net     0%   100  100  104.29  106.59  121.48
ge-1-0-0.418.rtr.chic.net.internet2.edu    1%    99  100  106.18  108.10  124.67
s4-ge-peer1.manlan-abilene.twaren.net     1%    99  100  126.76  130.56  246.96
s4-ge-chti-ir1.tpc-chi.twaren.net         0%   100  100  317.08  317.45  322.97
ascc-76-v20.twaren.net                    1%    99  100  317.57  318.76  381.60
202.169.174.90                            1%    99  100  315.55  317.39  438.84
202.169.174.50                            6%    94  100  315.57  316.50  341.02
202.169.174.29                            1%    99  100  315.54  316.77  353.03
202.169.175.101                           1%    99  100  316.03  316.81  323.32
140.109.98.111                            1%    99  100  315.50  316.32  323.82

Influence of TCP window size and number of parallel streams

Using the BIC congestion algorithm

From the TCP Tuning Guide:

Starting in Linux 2.6.7 (and back-ported to 2.4.27), BIC TCP is part of the kernel, and enabled by default. ... Starting with version 2.6.13, Linux supports pluggable congestion control algorithms

So, the default algorithm for this kernel should be BIC. This setting should only affect the sending side.

Upload to ASGC:

pIL_BIC-20070505-TO-140.109.98.111.png

Download from ASGC:

pIL_BIC-20070505-FROM-140.109.98.111.png

Using the Westwood congestion algorithm

TCP Westwood algorithm was activated by

sysctl -w net.ipv4.tcp_westwood=1

This setting should only affect the sending side.

As can be seen from the upload graphs, westwood shows no improvement over BIC for this connection. It even seems to perform slightly worse.

Upload to ASGC:

pIL_westw-20070505-TO-140.109.98.111.png

Download from ASGC:

pIL_westw-20070505-FROM-140.109.98.111.png

-- DerekFeichtinger - 05 May 2007

Edit | Attach | Watch | Print version | History: r2 < r1 | Backlinks | Raw View | Raw edit | More topic actions
Topic revision: r2 - 2007-05-05 - DerekFeichtinger
 
This site is powered by the TWiki collaboration platform Powered by Perl This site is powered by the TWiki collaboration platformCopyright © 2008-2024 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback