INN performances [Re: freenix stats with inn-cvs ... ?]

[Fabien Tassin]

According to Russ Allbery:
> 
> Fabien Tassin <fta at sofaraway.org> writes:
> 
> > Full feeds are about 1.2 to 1.7M articles these days so I assume the
> > limit should be much higher now but this has some bad effects.. smaller
> > sites become useless (ie irrelevants) for the top and it is also very
> > easy to cheat by sending billions of forged stats.
> 
> It also makes for interesting biases.  I care mostly about who's
> propagating the text discussion newsgroups, and quite a lot of those 1.7M
> articles are either binaries or jobs spew.

and spams plus the corresponding cancels :)

> A no-binaries top 1000 would be really interesting to me.

sure but by irrelevant, I meant that their data will have no influence
on the top. I've said it's a bad thing.
Propagating these 1.7M articles is not easy either. It requires a lot
of BW, a well tuned server and good peers. INN+innfeed still has some
problems to catch several full feeds at once. It is very common to
see two INN trying to push a full feed to the other side but spooling
at hell while both end-to-end BW and CPU/RAM not overloaded.

On that point, I'd like to understand why but until now, I'm puzzled.
I've added a timer in innfeed to see that I'm +92% idle and the rest
is just copying data again and again. My innd is also +60% idle and
is spending more time to write history than to write articles :(
Even with that, some of my peers have tons of backlog for me, almost
as much as I have for them, even if the network between us is very good
(plenty of BW available, no drops, short paths, etc.). INN tends to
be better as time passes but it still needs some major changes to
be a good transiter. I'm interested in opinion on that point.

-- 
Fabien Tassin -+- fta at sofaraway.org