Quoting Daniel Senie ([EMAIL PROTECTED]):
> At 02:20 PM 12/25/01, Chuck Yerkes wrote:
> >30 seconds is almost always a wait for a DNS lookup.
> 
> I've found that not to be the case. A 30 second delay will ONLY be DNS the 
> first time. The DNS server near (or co-resident) with the server providing 
> caching will cache either the positive or negative result. So, you might 
> see a 30 second delay once, but if you then check your mail a second time 
> and get another 30 second delay, that would point elsewhere.

Except when that DNS lookup fails.  Most DNS servers don't do
negative caching (storing failures).  When a connection comes
in, a lookup of the connecting IP address is done.  Unfortunately,
too much of the Internet is not run by the competent any more,
and those lookups fail.  My favorites are that 2.1.220.in-addr.arpa
(e.g.) will have NS servers that are unreachable.  This will cause
a 30 second delay each time.

> I've found the 30 second delay to be a telltale sign of IDENT processing in 
> XINETD. Indeed, in Xinetd, someone hard-coded 30 seconds as a timeout, and 
> didn't even provide a way to alter that value. Removing USERID logging in 
> your xinetd config will turn this off, give or take bugs in xinetd that 
> seem to turn it back on.

Dunno, don't really use Linux much.  When I do, I build from
source (a BSD habit).  I happily change the timeouts (it's the
point of using Open Source tools).

To make this note actually *useful*, I'll suggest that when
facing this sort of problem, a useful tool to debug is to run
tcpdump (or a tool on top of that) on the segment, filtering
for the connecting network or host.  You'll see that an IDENT
packet has gone out and, due to misconfigured firewalls, has
been dropped rather than rejected, causing delay.  (we found
THAT with sendmail a lot, leading to the current sendmail
default timeout of 2s rather than 30s for ident packets).

Sniff the line, see what's going on.  Run in debug mode and
watch.  Remove portions of the system (inetd, qpopper, use
different clients) and see where the delay is introduced.

Reply via email to