I've set up a hacked squid on each of the apaches that returns ICP_HIT on
any icp request. The squids are running on the apache machines at nice 19,
taking up <1% cpu. The apache to receive the next request is now
determined by the icp response time, basically the first icp reply from
the apache machines gets the request. High load -> slow icp response -> no
request this time.
Load distribution is more even than with the round-robin setup, especially
if something else is running on one of the apaches (or if the apaches are
not identical machines).
--
Gabriel Wicke