|
Chris Williams wrote:
>
> > keepalive is for maintaining an idle connection. The 2 minute timeout
> > is the FIN_WAIT for an InActConn to clear.
> The the thing is I'm only serving http, a connectionless protocol
http is stateless, but unlike some other stateless protocols (nfs, ntp)
requires a connection (ie uses tcp, rather than udp). http normally
disconnects quickly
http://www.linuxvirtualserver.org/Joseph.Mack/HOWTO/LVS-HOWTO-9.html#ss9.19
so that you'll not normally see any entries in ActConn and from that point
of view, the connection is short and looks a little like udp.
However for http/1.1 the connection can be persistent
(in the netscape sense, rather than persistent in the LVS sense)
http://www.linuxvirtualserver.org/Joseph.Mack/HOWTO/LVS-HOWTO-9.html#ss9.20
The persistence is only for 15secs, which is still relatively short, but
you'll still have to wait for the FIN_WAIT period of about 2mins before the
InActConn entry will be cleared from the ipvsadm table.
> Would this still apply? ipvsadm shows no inactive or active connections
> (once the initial download has occured,) but the problem still occurs
> that once I pull a real server out of the cluster, ipvsadm shows it as
> being removed, but traffic still gets directed there for 2-5 minutes.
hmm, once the entries have cleared from ipvsadm (seen in ActConn and InActConn),
you're free to remove the realserver. Are you using ldirectord (which
I don't know about) or something else that I don't understand here?
> Also I've just discovered that connecting from a different machine then
> the one doing load testing (a router unfortunetly, so not all together
> independant,) still gets sent to the dead real server.
something else is going on here. You haven't fiddled any timeouts have you?
> I'm starting to think that maybe I can live with a problem 2 minutes on
> server failure, but will struggle on for the moment.
you'll have to live with the FIN_WAIT timeout, but this isn't what's going on
here.
Joe
--
Joseph Mack PhD, Senior Systems Engineer, SAIC contractor
to the National Environmental Supercomputer Center,
mailto:mack.joseph@xxxxxxx ph# 919-541-0007, RTP, NC, USA
|
