Category :Network

Network Testing

To ensure maximum network performance, we ran several tests. First only with a cable running between two nodes then with one or many switches inbetween.

Back-to-back tests

Set-up

In this scenario we test the UDP and TCP performance between two computers with varying paramters. Usually we use different MTU sizes to see the effect of Jumbo frames, but many other parameters are also tested. Usually we use a script like this to run these tests.

ASCII-Art:

 +---------------------+                       +---------------------+
 | eth0: 192.168.0.202 | connection under test | eth0: 192.168.0.203 |
 | eth1: 172.20.1.202  +-----------------------+ eth1: 172.20.1.203  |
 +---------------------+                       +---------------------+

Here are some info of these systems: {| class="wikitable" border="1" |- ! type ! node 202 ! node 203 |- | sysctl | look at it | look at it |- | /proc/config.gz | look at it | look at it |- | ethtool | look at it | look at it |- |}

Kernel is 2.6.23.12, modinfo of e1000 tells me 7.3.20-k2 loaded with default settings.

Results

• First results are done with "pristine" netperf-2.4.4 (from Debian) without any special options can be found on this page.
• Second round is done with suggested parameters like -c -C and --use burst at compile time, these results can be found here.
• Third round, same settings as in round 2, but with tso off.
• Next round, irqbalance user space daemon running, but virtually no change.