First:
> Do other sites show the problem (eg try www.cymru.net for speed - that
> isnt using any kind of DNS/ident). What is the link MTU ?
You're right, this site is always fast. The only site I've had
troubles with are sites with lots of ads, like the Dilbert site I was
testing with.
Second:
Our site is serving our domain name over a 28.8k link. This would
appear to be causing some of the huge pauses when we have 75 or so
lines of the following in the tcpdump of the ppp link (we're
mango.mymenus.com):
00:36:36.595082 161.97.204.100.domain > mango.mymenus.com.domain: 31 A? mymenus.com. (29)
00:36:36.595082 mango.mymenus.com.domain > 161.97.204.100.domain: 31* 1/3/3 (164)
00:36:36.595082 mango.mymenus.com.domain > i.root-servers.net.domain: 48435 (45)
00:36:37.315082 i.root-servers.net.domain > mango.mymenus.com.domain: 48435- 0/3/2 (176)
00:36:37.315082 mango.mymenus.com.domain > E.ROOT-SERVERS.NET.domain: 48436 A? BVSD.K12.CO.US. (32)
00:36:37.315082 mango.mymenus.com.domain > ubu.Colorado.EDU.domain: 48437+ (45)
00:36:37.635082 E.ROOT-SERVERS.NET.domain > mango.mymenus.com.domain: 48436- 0/6/8 (312)
00:36:37.635082 mango.mymenus.com.domain > excalibur.usc.edu.domain: 48438 A? BVSD.K12.CO.US. (32)
...
Third:
I clearly have times when one machine can access the web page
'normally' (i.e., without pausing, loading at 3k/sec, etc.), while
another machine on our lan, loading the same pages, loads at 30
bytes/sec and stalls constantly. Earlier in the day 2 Linux boxes
were the stallers and 3 Macs were lickety-split, but now, at 5:00am,
my Linux machines are going quickly and the Macs are slow. It's
probably a red herring, because how could this be true, but it's been
a pretty convincing phenomenon (behaves consistently on five machines
over the course of a half hour).
Fourth:
The stalling isn't always associated with the dns domain transfers
shown above; sometimes it just appears as if the target site is
stalled. It probably is, though when it stalls from one machine and
not another, well, I don't know. When the stalling isn't from a dns
transfer, it usually follows one of these:
03:25:31.780611 0:0:c0:fd:ff:3 null > 0:a0:24:54:2d:7a sap 45 I (s=22,r=0,C) len=42
423e 0000 3f06 8522 ce9b 5e02 cf79 b854
3797 0050 31ea 76cf 0000 0000 6002 0200
011a 0000 0204 05b4 0d00 c000 0f00 c000
0200
Not knowing what that is, I thought it might be an appletalk thing, so
I compiled without appletalk, which made no difference at all. This
is probably another red herring.
I don't think it's my kernel version (I've tried a lot of them), and I
don't think it's the PPP parameters (I'm running without compression,
asyncmap 0, mru and mtu of 296 (and tried 542 and 1500)). I'm not
using NIS; tcpdump didn't indicate lan traffic causing the problems.
Thank you to everyone who has sent me ideas,
-- Pete Harlan pete@mymenus.com