CURRENT branch and older OSes
Russ Allbery
rra at stanford.edu
Mon Mar 23 01:23:55 UTC 2009
Julien ÉLIE <julien at trigofacile.com> writes:
>> Normally, Nagle's algorithm is a performance optimization that we want
>> (think collecting the pieces of an OVER reply into one packet).
> HDR, OVER and XPAT use stdio and such little pieces are sent with
> PushIOb only when they reach the size of BIG_BUFFER, as you once
> mentioned. So it should not affect these replies. The only replies it
> will affect are LIST replies (Printf on each line) but I did not notice
> a slow on Debian. With Wireshark, I saw that there was not a packet for
> each line but that they were sent grouped, as though TCP_NODELAY were
> overriden by another thing elsewhere.
You're right, I'd forgotten that we'd already done that work. Hm. Maybe
we should just enable it unconditionally, then, if it still makes a
difference. I suspect that it may change based on the TCP behavior of the
client, since the interaction with TCP delayed acknowledgements would be
with whether the client is delaying acknowledgements.
--
Russ Allbery (rra at stanford.edu) <http://www.eyrie.org/~eagle/>
Please send questions to the list rather than mailing me directly.
<http://www.eyrie.org/~eagle/faqs/questions.html> explains why.
More information about the inn-workers
mailing list