Ah, I completely forgot that we have 5 min. of persistence for
session-based web applications. That factors into the mix, but it
doesn't account for the extended swamping.
We use NAT, so the connection tracking/mapping might be part of the
trouble. I've had problems with incorrect SNAT from the real servers,
so there exists some degree of disconnect between ip_vs and netfilter.
I'll collect specific info the next time I take a server out of
rotation.
Here's another question: at what point in the connection negotiation
does lvs establish a connection? After the SYN/ACK? After the ACK? In
what state would the real server's application have to be for the port
to remain open, yet lvs not register new connections? Is that even
possible?
Thanks,
-jrr
On Wed, 2004-02-04 at 14:42, Joseph Mack wrote:
> so what happens to this server? For the next hour the customers are
> only getting service via the one server? Why doesn't everything equilibrate
> in 2 mins (TIME_WAIT)?
|