[nsp] L3 Loadbalancing EtherChannels

Oliver Boehmer (oboehmer) oboehmer at cisco.com
Wed Dec 24 10:17:37 EST 2003


> I can't seem to get the load balancing optimal between 2 links in an
> EtherChannel the way I want. This is the situation:
> 
> 1 Linux Router is connected to a Catalyst 4006 running CatOS
> 1 Server connected to the switch.
> 2 Cisco Border Routers are connected to the same switch.
> 1 of these Border Routers is connected with a 2 Gbit EtherChannel.
> 
> The Linux router is talking BGP to both Border routers to decide
> where to send it's traffic.
> 
> The catalyst shows a 52%-48% packet balancing however the traffic
> distribution is 300Mbps for 1 link and 100Mbps for the other one. As
> far as I can see traffic from the server is using the 100Mbps link,
> all traffic from the Linux router is using the other link. The Linux
> router is routing NNTP traffic with has a packet size of 1500+ bytes.
> The other link has much smaller packets which explains the 52-48
> packet balancing and the 300-100 traffic utilization.
> 
> The c4006 is not running IOS you can't chance the distribution method.
> 
> Since I can't get the link utilization correct on L2 I'm looking for
> ways to do so on L3. Can anyone give me some suggestions? CEF based
> load balancing doesn't work since Quagga (Linux) doesn't support it...

Hmm, as a "hack" you could configure multiple HSRP groups on your Border
router's PortChannel interface and use outbound route-maps to modify the
BGP next-hop to the alternate virtual HSRP IP addresses:

int portchannel1
 ip address 192.168.1.254 255.255.255.0
 standby 1 ip 192.168.1.253
 standby 1 priority 255
 ! high priority just to make sure we'll be master..
 standby 2 ip 192.168.1.252
 standby 2 priority 255
 standby 3 ip 192.168.1.251
 standby 3 priority 255
 
As each virtual IP address uses a different MAC address, you will
effectivly create multiple <source-mac,dest-mac> L2 flows resulting in a
more optimal distribution as long as the number of prefixes is large.. 

Multiple equal cost default routes to the various virtual addresses
should achieve the same on the server (if the server supports
load-sharing across equal paths).

	oli



More information about the cisco-nsp mailing list