Statistical Variations as a function of Ping Number

Aim

To determine whether the number of packets used to sample affects the statistical results obtained.

Method

This was conducted on the Dell Dual Xeon 1.7 machines on the GigE interface, with minimal load on the machines.

Using constant intervals of 1 second and a packet size of 1400 (ICMP, ie 1408bytes IP) for every packet sent, we varied the number of packets sent. The min, max, ave and stdev of the sample was then found.

In order to overcome arp recaching, a single ping of the same type was sent before the test for each test.

A rather large sample set was used ranging from 1 single ping to 100 pings.

Scripts

A wrapper script around ping was used.

do_number.pl - given a list of numbers, will conduct the ping tests and write a summary table of results.
ping-cook.pl - required script to generate table information for ping

Results

The complete set of log files can be found here.

There was no packet loss and no reordering of the packets

Surprisingly, there is a slight variation in the minimum rtt. This is most likely due to cpu load on the source and or sink causing kernel space time differences in processing the echo/reply. It may also be due to interrupt coalesing.

The average rtt is pretty constant (ignore the first point - slight bug in the calculation of ave for single values - it's been sorted), with a an average of 0.201 (3dp) from the average of the averages. From this run, if we were to use less than 3 packets, we get a slightly higher value of about 0.203/5.

Wed, 23 July, 2003 13:07