Comparision of iperf statistics and web100 statistics for the same connection

AIM

This experiment is to show the accuracy of iperf statistics. (of course, this assumes that web100 is correct). Special consideration of the 4gb limit in iperf is explored and the results presented.

Method

This experiment is composed of two runs. The first run investigates the accuracy of the amount of data that is transferred. We will do this by using the -n option in iperf and comparing the amount of data transfered by iperf with that reported in web100. We will also look into the reported values of throughput for these various values by dividing it by the duration of the experiment.

The second part of this experiment will investigate the reported values of the throughput when iperf transfers more than 4gb. The only way this is possible is through defining a duration in which the test should be sent.

Run 1

conducted on 9th July 2002.

By varying the amount of data sent into the network with iperf, we passively monitor the amount of data send into the network for various data sizes.

The data sizes have been decided to be {64,128,256,512,1024,2048,3072,4096}mbytes. As we can get about 900mbit/sec, this is about 112mbytes/sec. So for the largest transfer size of 4gb, it should take about 35 seconds. We need to set web100 trap rates, this was chosen to be 10ms. This would generate a lot of information.

A future tests will investigate into the performance affects of monitoring with web100.

This will be repeated three times to make sure the results are okay. Each test will have the reciever (pc56) set with a buffer of 2048m and the sender with a socket buffer size of 2048m also.

The web100 variable to compare against are:

Data Size: DataBytesOut
Duration: CurrTime - Delta
Recievers window: CurrentRwinRcvd
Socket buffer size on sender: Sndbuf
Socket buffer size on reciever: Rcvbuf

Test 1

results here.

Test 2

results here.

Test 3

results here.

summarising these results:

It seems like on average, the reported values (iperf raw) is, on average, higher than that reported by iperf by about 0.5% - with quite a high deviation. The difference between the raw iperf value and the calculated can be attributed to the inaccurate measure/display of the test duration - although from the above results, to one decimal place, the reported duration is accurate. There also seems to be some discripency regarding the actual amonut of data that is being transfered - with a 4gb transfer sending out (according to web100) about 500 bytes more than specified. It is worth noting that this is actually a very small amount.

However, it does show that the results between the recorded iperf results and results collated with web100 are not wholly consistent, and that iperf can seem to over predict the value of path utilisatio by upto 30 out of 900mb/sec, ie about 3%.

I am currently running iperf tests with web100 passive monitoring on GridNM.

Run 2

For various test durations that exceed sending 4gb of data, we monitor the amount of data sent into the network with web100.

Wed, 23 July, 2003 13:07