Re: [nm-wg] treatment of lost packets when measuring delay

[stanislav shalunov <shalunov@internet2.edu>]

Augusto Ciuffoletti <augusto@di.unipi.it> writes:

> [...] I came to two conclusions [...:]

> -) statistical figures are of little help, unless they are in some
> way filtered to take into account that recent values are more
> significant then older ones: while it is quite clear what is a
> "weighted average" is, it is not as straightforward to define a
> weighted 95-ile. Anyway, it is better to be explicit about whether
> the statistic is filtered or not;

What parameter of what distribution is being estimated by computing a
weighted average of delays below a given threshold?  Certainly not the
mean of the underlying delay distribution.

> -) a weighted statistic is less prone to long range dependency than
> a long term one (but we exclude that the user will use tomorrow the
> statistics collected today);

I assume you mean weighted average.  Weighted average is not a robust
estimator: a change in any one value in the sample can significantly
change the estimate.

> -) losses and delays are (or should be) symptoms of distinct events:
> losses indicate a problem, delays quantify expected operation. The
> fact that they can be put on the same axis (the "delay") is, imho,
> misleading: so I'd advice separate figures, both approrpiately
> filtered. This implicitely solves the hairy problem of including an
> infinite value in a sum :-)

What is externally observed loss, if not non-arrival before a timeout?

And what, again, is the problem with computing a sum where a few
values are +\infty?  (Note that we can't find an actual infinite delay
by external observation---we can only make sure that a delay is
greater than some value.)

While this is not central to the discussion, losses certainly do not
always indicate a problem: in addition loss is used to communicate
from the routers to the senders.  Whether we like it or not, such
communication is the standard way of running TCP and a whole research
area clusters around active queue management---which, in essence, is
all about ways to use loss to communicate more efficiently.

-- 
Stanislav Shalunov		http://www.internet2.edu/~shalunov/

This message is designed to be viewed with 0.06479891g of NaCl.