LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: Easing a real server into rotation

To: "LinuxVirtualServer.org users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: Easing a real server into rotation
From: Horms <horms@xxxxxxxxxxxx>
Date: Mon, 9 Feb 2004 09:41:23 +0900
On Fri, Feb 06, 2004 at 10:07:24AM -0500, John Reuning wrote:
> On Thu, 2004-02-05 at 21:24, Horms wrote:
> 
> > I suspect that your web browser (Explorer?) is reopening the connection
> > (RFC 1122, 4.2.2.13), but that it is very short lived so you can't
> > observe what is going on. So basically the port 49598 connection is
> > swiching betweeen active and inactive, which is why you don't see the
> > inactive connection count grow.
> 
> I was using Mozilla.  However, shouldn't telnetting to VIP port 80 cause
> the connection to relist as ESTABLISHED until the telnet session is
> ended?  I tried that, and nothing changed.
> 
> And I noticed something else, too.  After the connection disappeared
> from the table (after the TIME_WAIT expired), it didn't reappear at all
> when I reloaded pages via a browser or telnetted to port 80.  Only the
> persistence SYN_RECV line remained.  It wasn't until after the SYN_RECV
> line timed out that a new ESTABLISHED connection line appeared in the
> table.

Are you sure you are connecting to the linux director and not
mistakenly connecting directly to one of the real servers?
Are you using LVS-DR? If you are and you have not hidden
the lo:0 interfaces on the real servers then this
is a very real possibility. Alternatively, perhaps
something strange is going on with ICMP redirects. Have you
traced what is going on with tcpdump?

> > I am not sure if this helps or not, but if a connection is in the
> > connection table (i.e. it shows up in ipvsadm -L -c -n), but is not a
> > persistance template (i.e. the client's port is not zero) then it will
> > be counted in either ActiveConn or InActConn. If the connection is in
> > the ESTABLISHED state then it is counted in ActiveConn, otherwise it is
> > counted in InActConn.
> 
> Yep, this makes sense.  And it leads me to think that the root cause of
> the newly added real server may be the result of inaccurate ActiveConn
> counting.

I haven't seen anything in what you have reported to indicate
that the connection counts are inaccurate.

-- 
Horms
<Prev in Thread] Current Thread [Next in Thread>