[squid-users] 3.3.x -> 3.4.x: huge performance regression

[Eugene M. Zheganin]

Hi.

I was using the 3.4.x branch for quite some time, it was working just
fine on small installations.
Yesterday I upgraded my largest cache installation from 3.3.13 to 3.4.8
(same config, diskd, NTLM/GSS-SPNEGO auth helpers, external helpers).
Today morning I noticed that squid is spiking to 100% of CPU and almost
isn't serving any traffic. Restart didn't help, squid is serving pages
while continuing to consume CPU, load grows, until it's at 100%, and
after some time my users are unable to open any page from Internet. This
is sad, so I downgraded to 3.3.13. CPU consumption went back to 20-35%
and everything is back to normal.

In order to understand what's happening I did some dtrace profiling to
see what is squid busy with, taking the consideration, that measuring
the same amount of connect()/socket() syscalls should give same amount
of squid work, but the results were totally different on one number of
such syscalls.

Anyone to comment ?

Thanks.
Eugene.