[lvs-users] Connections are "clumped" together on real_server failure

Malcolm Turnbull malcolm at loadbalancer.org
Tue Jun 30 10:37:49 BST 2009


Jay,

Try switching from WLC to WRR? (I'm guessing that you are using WLC).



2009/6/25 Jay Faulkner <jay.faulkner at mailtrust.com>:
> Hey guys,
>
> We've been noticing some really strange behavior with keepalived/LVS when a real_server fails. The behavior we see is that when a r_s fails, all the persistent connections on that server fall over to a single server.
>
> For example:
>
> Server 1 - 25000 active conn. - 1500 inactive conn.
> Server 2 - 35000 active conn. - 1000 inactive conn.
> Server 3 - 20000 active conn. - 500 inactive conn.
>
> Server 2 fails. Instead of seeing all the traffic split between the remaining real_servers, we see it all clump over to the same server
>
> Server 1 - 25000 active conn. - 1500 inactive conn.
> Server 2 - failed
> Server 3 - 55000 active conn. - 1500 inactive conn.
>
> Now, server 3 falls over and dies due to having too many connections on it and becoming overloaded.
>
> We've seen this behavior consistently in all of our keepalived installations (20 servers or so? Median of about 20 virtual_servers each), and I was curious if this is configurable behavior, or even expected.
>
> Thanks!
>
> Jason Faulkner
> Linux Engineer, Rackspace Email & Apps
> jason.faulkner at rackspace.com<mailto:jason.faulkner at rackspace.com>
> o: (540) 443-2101 (ex. 505-2101)
>
> _______________________________________________
> Please read the documentation before posting - it's available at:
> http://www.linuxvirtualserver.org/
>
> LinuxVirtualServer.org mailing list - lvs-users at LinuxVirtualServer.org
> Send requests to lvs-users-request at LinuxVirtualServer.org
> or go to http://lists.graemef.net/mailman/listinfo/lvs-users
>



-- 
Regards,

Malcolm Turnbull.

Loadbalancer.org Ltd.
Phone: +44 (0)870 443 8779
http://www.loadbalancer.org/




More information about the lvs-users mailing list