TXQueueLen Investigation into IP Performance

Background

With the advent of high speed communications, larger queues are required to store all the extra data that is shoved onto network. Apart from the way in which different protocols above Layer 2 encapsulate and regulate the way in which the data is sent out (to prevent congestion collapse), is the way in which Layer 2 devices such as Ethernet NICs put the packets onto the 'ether'.

In Linux Kernels, there is a setting available to regulate the size of the queue between the IP layer and the Ethernet layer. Careful tuning is required to ensure that the size of the queue is optimal for your network connection.

This report investigates the effect of varying the txqueuelen using the ifconfig parameter and studying the effects of udp and tcp traffic at Gig speeds.

Documents

Queues in the Linux kernel by Mathieu Goutelle

Method

This report will be structured into 4 parts:

• B2B udp test results
• B2B tcp test results
• MBNG udp test results
• MBNG tcp test results
• DataTAG tcp test results

Each test will be conducted for single stream between Dual Pentium Hyperthreaded enabled machines with e1000 4.4.12 drivers with default settings. The Linux kernel under investigation is 2.4.19 with web100 version 2.1/2.2 for tcp tests.

B2B tests simply mean a couple of the high end pc's connected with a CAT5e cross cable cable. MBNG imply tests conducted through the MB-NG test network with network latency of 6ms. The complete end to end path consists of 3 Cisco 12000GSRs and 2 Cisco 7600s. The PC on the other end is identical to that of the sender.

Wed, 23 July, 2003 14:01