LVS
lvs-users
Google
 
Web LinuxVirtualServer.org

Re: [lvs-users] LVS with all clients behind a single router. lvs-tun on

To: "LinuxVirtualServer.org users mailing list." <lvs-users@xxxxxxxxxxxxxxxxxxxxxx>
Subject: Re: [lvs-users] LVS with all clients behind a single router. lvs-tun on 2.6 kernel
From: Graeme Fowler <graeme@xxxxxxxxxxx>
Date: Sat, 23 Jun 2007 20:31:47 +0100
On Sat, 2007-06-23 at 13:30 -0500, Matthew Smart wrote:
<snip LVS with multiple NAT client explanation>

> The problem I have is that all clients from a given location get 
> directed to the same realserver. Since the majority of clients are 
> located in the same office, we are not getting a good load balance. I 
> found the following thread in the archives which addresses this issue:
> http://archive.linuxvirtualserver.org/html/lvs-users/2005-04/msg00116.html
> 
> That thread mentions that the realservers need tuning for ARP. I have 
> used the arp flags for 2.6 in the manner show in howtos online. Can you 
> elaborate on this comment and possibly direct me to some resources 
> discussing what can be done in this circumstance?

I don't think you have an ARP problem - your LVS is working as designed,
if not as desired.

I'm not sure, given that you only ever have three client addresses here,
that you'll get the behaviour you want - it'll always be imbalanced with
such a low number of client addresses.

*However* - you could get creative - well, I think this would help you,
anyway...

The clients will only come from three source IP addresses, but
presumably if they have half-decent NAT kit that means a range of source
port numbers will be involved (you'd have to work the ranges out by
observation, I expect).
You could then setup a bunch of inbound netfilter (iptables) rules on
the director(s) to assign a different netfilter MARK value (fwmark
value) to various source port ranges; you could then have multiple
virtual services assigned via fwmark values instead of VIP (you'd still
need the VIP on the realservers for this to work though, if you're using
DR).

Just as a theoretical example, using three realservers:

group source ports 1-20000 connecting to VIP into fwmark 1
group source ports 20001-40000 connecting to VIP into fwmark 2
group source ports 40001-65535 connecting to VIP into fwmark 3
(you could make this as granular as you want to)

have a virtual server definition for fwmark 1:
  RS1 weight 200
  RS2 weight 100
  RS3 weight 100

have a virtual server definition for fwmark 2:
  RS1 weight 100
  RS2 weight 200
  RS3 weight 100

have a virtual server definition for fwmark 3:
  RS1 weight 100
  RS2 weight 100
  RS3 weight 200

The total weight for each server is identical (400) but the spread
within each group is slightly different. The clients, however, should
see no change.

The only problem I can see with this is that persistence will only work,
I think, if the clients stay within the same port range. If a given
client's source port drifts into a different range then it will, very
likely, drift onto a different RS - but persistence on all three virtual
services might take care of that, given that persistence is actually
done by source IP; as long as all three realservers are included in each
virtual service I *think* the persistence will carry. Horms or Joe might
want to check that statement, mind you.

Give it a try and see if it helps.

Graeme



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