[squid-users] getting contents from sibblings

Hi,
We have a squid-cache cluster that is supposed to be configured
as an accelerator for multiple websites.
We use icp to keep the caches in sync. But lately we noticed the
caches do not co-operate anymore. A test setup of (almost) the
same configuration still works.

Here is an example of a working request (http obfuscated to xyz):
After doing this wget a few times I get this:

ard@c24574:/net/home/ard$ wget --proxy=off -SHO /dev/null 'xyz://test.dft.nl/servlet/ArticleByID?ID=4158800&section=00'
--13:36:41-- xyz://test.dft.nl/servlet/ArticleByID?ID=4158800&section=00
           => `/dev/null'
Resolving test.dft.nl... done.
Connecting to test.dft.nl[217.196.32.34]:80... connected.
HTTP request sent, awaiting response...
 1 HTTP/1.0 200 OK
 2 Date: Wed, 19 Feb 2003 12:32:56 GMT
 3 Server: Apache/1.3.26 (Unix) Debian GNU/Linux Resin/1.2.8
 4 Last-Modified: Wed, 19 Feb 2003 12:32:56 GMT
 5 Cache-Control: max-age=600
 6 Expires: Wed, 19 Feb 2003 12:40:00 GMT
 7 Content-Length: 31216
 8 Content-Type: text/html; charset=iso-8859-1
 9 X-Cache: HIT from squid2-accel-bart
10 X-Cache: HIT from squid1-accel-seabert
11 X-Cache: HIT from squid1-accel-henk
12 Age: 225
13 X-Cache: HIT from squid1-accel-midas
14 Connection: keep-alive

100%[===========================================================================================>] 31,216 29.77M/s ETA 00:00

13:36:42 (29.77 MB/s) - `/dev/null' saved [31216/31216]

This clearly shows that bart got it first, then seabert got it from
bart, then henk got it from seabert, then midas got it from henk, and
we got it from midas.

Doing it again later (about 20 minutes after the expire) none of the
caches takes the trouble of requesting it from a neighbour cache.

It seems that squid somehow prefers the real servers instead of the
sibbling caches. How does squid determine if a sibbling should be used
instead of the real server. Is there any clear documentation about
that?

More information about the setup:
There is a farm of servers behind a round-robin load-balancer, with
each server running 1 squid per cpu (2 squids per system)

icp_port 3130
icp_access allow icpdomain
cache_peer robin sibling 80 3130
cache_peer robin sibling 81 3131
cache_peer rakker sibling 80 3130
cache_peer rakker sibling 81 3131
cache_peer henk sibling 80 3130
cache_peer henk sibling 81 3131
cache_peer midas sibling 80 3130
cache_peer midas sibling 81 3131

icp_query_timeout 60

I know this is not enough information, but I would rather like to know
where to look, and then give the relevant information...

Regards,
Ard

-- 
program signature;
begin  { telegraaf.com
} writeln("<ard@telegraafnet.nl> SMA-IS | Geeks don't get viruses");
end
.