Re: [bmwg] Mean vs Median

[Paul Emmerich <[e-mail]>]

Hi,

On 10.11.15 02:18, Marius Georgescu wrote:
>  > Can you give us more context (test setup; physical/virtualized
> tester/DUT; one tester/sender_receiver tester ... ) on these measurements?

I also found an example for a non-normally distributed latency on a 
hardware switch.

Test setup: two flows of traffic coming from different ports on the 
switch that are directed to a single output port. The sum of two 
incoming flows is less than the capacity of the outgoing link, so this 
is not an overload scenario but a typical load scenario for a switch.

This setup results in a bimodal latency distribution, each packet either 
has to wait for a short time if there is a packet from the other flow 
currently being transmitted, or it can be transmitted immediately if the 
link is free.

The buffers on the switch are virtually empty as I used constant 
inter-packet spacings.


The histogram of a 1 Gbit/s flow is available at [1]. The second 
competing flow was 4 Gbit/s in this test setup, however, the slower flow 
shown in the graph was prioritized (which doesn't matter as the output 
was not overloaded). Changing the relative rates changes which of the 
two peaks dominates, but the basic building blocks are still the two 
normal distributions.

A median value is often completely meaningless in such a scenario.


Paul

[1] https://www.dropbox.com/s/xbgelt336ade0dx/two-flows-pica8.pdf?dl=1