[Netalyzr] upstream buffering test

[Nicholas Weaver]

On Feb 26, 2012, at 9:25 PM, Brian J. Murrell wrote:

> Hi,
> 
> I have tried running your netalyzer on my network here and I want to
> verify the upstream buffering results.
> 
> I run QoS here on my network and see quite stable ping times while I
> saturate my upstream bandwidth.  During such a test I am doing:
> 
> $ scp a_big_file $host
> 
> and while doing that I run "mtr $host" to watch the jitter of each hop
> along the path.

Excellent!

> 
> $host is 9 hops away and usually about 25ms +-3-5ms.  While I'm scp'ing
> that big file (about 17MB) the standard deviation for most hops on the
> path to $host remains in the 5-10ms range which about the same whether
> the upstream is saturated or not.
> 
> So my question is, how is netalyzer testing my upstream buffering such
> that it's seeing a second or more of buffering when I'm not seeing it
> using the above test?

Because its a deliberately crude test that only works on FIFO buffers:  It consists of a full rate UDP flow (it basically is in continual exponential ramp-up) which measures the increase in latency during the last 5 seconds of the 10 second test.

If you did not have the active queueing, you would see 1s of latency, but since you have active queueing on the bottleneck buffer, this test reports the queue as being big (it is, for a single stream) but it doesn't affect the latency on your network under load.