UDP Flows - Varying Packet Size - 20th Feb 2003 ~ Midnight

Purpose and Method

The purpose of this test is to determine the performance of the MBNG for various packet sizes. This will be conducted by using iperf in udp mode and sending udp streams on an unloaded network. Eack packet size will be run for 60 seconds. As the iperf server need to know explicitly the packet size to accept, the server was reinitiated each run to the appropiate packet size.

In order to guarantee that iperf is sending at it's fastest rate, we set for every packet size the sending rate (-b) to 2000m.

The udp-packetsize.pl script was used which uses the iperf.pl script to initiate and shutdown servers and clients.

Experiment 1: mbgn1->gig1

log tsv zip

Experiment 1: mbgn2->gig5

Test A: 20th Feb 2003

log tsv zip

Test B: 21st Feb 2003

log tsv zip

Experiment 3: mbgn3->gig6

Test A: 20th Feb 2003

log tsv zip

Test B: 21th Feb 2003

log tsv zip

Experiment 2: Cross Traffic

To determine where the problems lie in terms of the packet loss and throughput degradations, it was decided to conduct the tests from UCL to the pcs at MAN which were not performed above. The results of this should show where the problem lies.

These tests were performed using a txq of 100 and standard 2.4.19 settings on a web100 2.1a kernel.

Test 1A: mbgn1->gig5 @ 21st Feb 2003 ~ Midnight

log tsv zip

Test 1B: mbgn1->gig6 @ 21st Feb 2003 ~ 00:16

log tsv zip

Test 2A: mbgn2->gig1 @ 20th Feb 2003 ~ 20:40

log tsv zip

Test 2B: mbgn2->gig6 @ 21st Feb 2003 ~ 22:56

log tsv zip

Test 3A: mbgn3->gig1 @ 20th Feb 2003 ~ 20:40

log tsv zip

Test 3B: mbgn3->gig5 @ 21st Feb 2003 ~ 22:56

log tsv zip

Summary

It appears that the problems of throughput may be associated with the hosts at Manchester. All tests from any machine to gig6, for example, has the same very low throughput for small packet sizes - only being okay at standard MTU. Similarly, all tests from any machine to gig5 appears to have okay throughput, except for low packet sizes where there appears to be loss. All tests to gig1 appears to be fine except for a mysterious region between 600 and 1000 where there are packet drops.

In terms of the iperf software, there appears to be a bug? in which it reports a rate much higher than it is possible to send. As a result loss occurs. Investigation into is required. Suggestions include:

• using a smaller value for the requested throughput
• look at the actual number of packets that comes out of the interface via the ifconfig util
• use another tool (udpmon, ipflow etc.)

Wed, 23 July, 2003 13:07