GoZilla vs. Squid

More information on my problem with Squid starting multiple parallel
transfers for the same URL:

I am running Squid 2.2.STABLE4-hno.19990807 on Linux kernel 2.0.36.
One of my users started a download of the URL

http://www.geocities.com/EnchantedForest/Pond/8851/6.doc

(a 5 MB MP3 file) with GoZilla. This ran for about 12 hours, at least
half of that time filling 128 kbps of bandwidth all by its own, and
still hadn't managed to get the file. During that time, I noticed in
the cachemgr.cgi filedescriptors page, each time I looked, several
(between 3 and 5) server side sockets for that URL open at the same
time to different IPs. (www.geocities.com has 11 IPs associated with
it in DNS.) Example:

   25 Socket 1439 299* 309 192.168.70.237.2786 http://www.geocities.com/EnchantedForest/Pond/8851/6.doc
   26 Socket 15 153196* 816 209.1.224.14.80 http://www.geocities.com/EnchantedForest/Pond/8851/6.doc
   32 Socket 14 1204458* 816 209.1.224.13.80 http://www.geocities.com/EnchantedForest/Pond/8851/6.doc
   39 Socket 14 3675756* 816 209.1.224.16.80 http://www.geocities.com/EnchantedForest/Pond/8851/6.doc

I never saw never more than one client connection (the 192.168.70.
address) though. In access.log, I found:

952508990.014 18215 PROMETHEUS.sema.de TCP_MISS/000 30369 GET http://www.geocities.com/EnchantedForest/Pond/8851/6.doc - DIRECT/209.1.224.18 -
952509032.245 3362 PROMETHEUS.sema.de TCP_MISS/000 1692 GET http://www.geocities.com/EnchantedForest/Pond/8851/6.doc - DIRECT/209.1.224.10 -
[... 139 similar lines deleted ...]
952550439.933 693386 PROMETHEUS.sema.de TCP_MISS/000 1719 GET http://www.geocities.com/EnchantedForest/Pond/8851/6.doc - DIRECT/209.1.224.14 -
952550775.464 696015 PROMETHEUS.sema.de TCP_MISS/000 2743 GET http://www.geocities.com/EnchantedForest/Pond/8851/6.doc - DIRECT/209.1.224.15 -

ie. lots of incomplete TCP_MISSes with varying server addresses and
no apparent progress. At that point I stopped the farce, for lack of
access to the client, by inserting an ACL forbidding that particular
transfer into my squid.conf.

The next morning, on my advice, the user tried downloading the same
URL with wget from different machine instead (through the same Squid)
and it came in without a hitch (well, almost):

952597843.076 4113 mastermind.sema.de TCP_MISS/000 2868 GET http://www.geocities.com/EnchantedForest/Pond/88551/6.doc - DIRECT/209.1.224.15 -
952597865.773 967 mastermind.sema.de TCP_MISS/000 2868 GET http://www.geocities.com/EnchantedForest/Pond/88551/6.doc - DIRECT/209.1.224.18 -
952598903.052 1033348 mastermind.sema.de TCP_MISS/200 4978313 GET http://www.geocities.com/EnchantedForest/Pond/8851/6.doc - DIRECT/209.1.224.12 application/msword

Unfortunately, this seems to happen every time somebody tries to
download some difficult to get URL (typically MP3s) with GoZilla.
Can anyone confirm such a detrimental interaction between Squid
and GoZilla, and/or suggest a configuration change to avoid it?

-- 
Tilman Schmidt          E-Mail: Tilman.Schmidt@sema.de (office)
Sema Group Koeln, Germany       tilman@schmidt.bn.uunet.de (private)