socket options & read raw

[Cyrill D. Schneider]

Andrew Tridgell wrote:
> 
> > solves several performance issues, obviously on all kinds of
> > different platforms. I've then sent a couple of logs to Jeremy Allison,
> > but didn't get a reply so far. Maybe they were completely useless...
> 
> More likely he is just very busy :)

I didn't expect an instant fix or whatever. I haven't even said it's a
bug at any point. But a quick reply telling me that he received the logs 
and will take care of the problem in 1999 or whatever would have been 
nice :)

> We get thousands of emails to samba-bugs and can only deal with some
> of them. Analysing performance logs takes a long time.
> You will also find that the settings you have discovered are only the
> best for some client/server combinations. In some cases setting
> TCP_NODELAY actually makes performance much worse. The same goes for
> "read raw=no".

I agree. But you got me wrong I think. This isn't really about tuning. 
Without these settings, the Samba server can be, basically, useless 
for quite a number of people out there. I hear from Linux, AIX and
Solaris
users and they all say that it changes throughput by a factor of 3-10
with
programs that ran dog slow before. Although the Samba box is now, in
general,
a bit slower with "read raw" disabled, it nevertheless makes a huge
difference 
with some (special) applications. And it's a tricky task to justify why
our 
MS Access applications run 5x slower on the Samba server than they did
before 
on NT, although it's the same hardware... In fact, most of them were
about
to trash the Samba server again, which would have been a shame IMHO.

Anyway, I never intended to offend anyone and I'm sorry if I did. Won't
bug
you again.

--Cyrill