On Wed, Aug 31, 2011 at 1:09 PM, Camaleón
<noelamac@gmail.com> wrote:
Hello,
I've been busy on these days trying to solve a problem with Postfix that
drove me nuts.
Sporadically (let's say one in hundred e-mails) my Postfix had problems
for delivering messages with ~3 MiB of attachment to some e-mail hosts.
DSN service returned the final notice of delivery to the user and logs
displayed an error like "timed out while sending message body".
These hosts were not of those "difficult" ones like Hotmail, Gmail,
Yahoo! or the like that because to their high volume of traffic implement
additional (and sometimes strambotic) measures to prevent spam and such
"anti-all" systems that may require a different transport definiton in
Postfix to get e-mails delivered.
Moreover, these hosts were not e-mail servers that are behind Cisco PIX
devices or using MS Exchange servers that are also well-known to be
conflictive to "dialogue" with.
Nope, I was having problems for delivering to common, small hosts of mid-
size companies, one of the hosts running a Debian system, like mine. So I
had to run some tests to find out what could be the problem here.
I first tried to define a less conservative values (by increasing the
time) for "smtp_data_done_timeout", "smtp_data_xfer_timeout" and
"smtp_data_init_timeout" but this had no effect at all and again, some e-
mails were still undelivered.
Googling around I found some posts and articles¹ pointing to the MTU
value (my bonded interface was set by default to 1500) and as I had
nothing to lose, I changed this and lowered to 1400.
This turned out to work wonders and since then (that's more than a week
ago) I still had no other DSN delivery errors. Besides, e-mails in
deferred queue that could not be sent in that time, after lowering the
MTU value were also delivered with no apparent problems.
I'm still monitoring this but if this is the "cure" to prevent such
errors, are there any expected drawbacks for lowering MTU "system-wide"?
The server has dual gigabit NIC which are bonded (in backup mode) and
server itself is behind a FTTH gigabit router. The server also hosts a
web server.
Any comments or experiences on this are welcome :-)
¹http://www.hsc.fr/ressources/cours/postfix/doc/faq.html#timeouts
You might want to try the postfix mailing list and see if they have any ideas.
Be prepared for cold, hard, terse answers. They don't chat much - busy I guess.
Was the previous MTU of 1500, a value you had set, or the default when queried?
I'm wondering because of a recent experience I had tweaking MTU. I wanted to try jumbo frames
to improve samba throughput on large video files. With MTU set to 9000 on Linux and Windows,
throughput increased about 8 times. But it caused problems with the web service on Linux,
which was running a domain under dyndns. I set the MTU on Linux back to unspecified, but
left the jumbo frame active on Windows side. The performance was still very good in large
samba file transfers. I might remember this wrong, but it seemed it was worse performance
when Linux side specified 1500, so I left it as unspecified and it has worked well.
I still get high transfer speeds in samba with unspecified MTU on Linux but jumbo
of 9000 MTU on Windows side.
BTW, 1492 is a common MTU seen in FAQs. You might get just as good with that.