LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: PATCH: synchronisation and active/inactive connection count

To: lvs-users@xxxxxxxxxxxxxxxxxxxxxx
Subject: Re: PATCH: synchronisation and active/inactive connection count
From: Horms <horms@xxxxxxxxxxxx>
Date: Mon, 1 Jul 2002 15:12:16 +0900
On Fri, Jun 28, 2002 at 10:23:01PM +0800, Wensong Zhang wrote:
> 
> 
> Hi Horms,
> 
> On Fri, 28 Jun 2002, Horms wrote:
> 
> > 
> > Interestingly I note that if a backup linux director takes over a
> > connection, its connection counters are not updated accordingly.
> 
> Yes, the synchronized connections don't affect the server connection 
> counters. After the backup takes over, new connections can be recorded in 
> the server connection counters. And, in the most cases, the synchronized 
> connections expires soon, so it doesn't affect the load balancing among 
> the servers too much.
> 
> > This is not a big problem, and doesn't cause any negative counts to
> > crop up, but may be worth fixing at some stage. What are your
> > thoughts?
> > 
> 
> There are several reasons that synchronized connections are not used in 
> the server connection counters.
> 
> 1) In many HA implementation, the backup load balancer doesn't have the 
> ipvs scheduling table when it is in the standby mode. So, when the backup 
> receives the synchronized connections, there is no server entry to update 
> connection counters.
> 
> 2) Even if the backup load balancer has the ipvs scheduling table, it is 
> hard to lookup its server entry (through the address and port numbers). 
> For example, you have to look up the virtual service first, if the 
> connection belongs to the fwmark-based services, it is not possible to get 
> the service entry through address and port. Then, you will have to 
> synchronize the connections with its service info, it will transfer more 
> data and bring more overhead.
> 
> 3) Keep it simple :)

Hi Wensong,

That makes a lot of sense. I guess I was just supprised
that the connections didn't show up. I think that given
that this only effects connections that are active when a 
fail-over takes palce it isn't really much of a problem
after all.

-- 
Horms
        


<Prev in Thread] Current Thread [Next in Thread>
  • Re: PATCH: synchronisation and active/inactive connection count, Horms <=