On Wed, Oct 08, 2008 at 09:14:02AM +0200, BARDOU Pierre wrote:
quoted text
>  
> Hello,
> 
> Failover already works with BGP on my test conf, the problem is that BGP 
> only selects ONE route to a destination, so there is no load balancing.
> 

There is loadbalancing insofar that if you have two independent upstreams
you get two different views of the internet and you should be able to
split the 250k IPv4 routes into two sets that will result in equal use of
both links.  This is the usual traffic engineering done on BGP with the
help of match filters that change the localpref based on communities, AS
pathes or whatever you like.

quoted text
> The easiest for me would be to tell BGP to keep TWO routes to each
> Destination, and use them in a round-robin way.
> 
> That's what Cisco does with BGP multipath
> http://www.cisco.com/en/US/tech/tk365/technologies_tech_note09186a0080094431
> .shtml#bgpmpath
> 

This will not work as you expect. In your setup case with two independet
upstreams only one upstream will be selected.

From the document:
In order to be candidates for multipath, paths to the same destination
need to have these characteristics equal to the best-path characteristics:
  * Weight
  * Local preference
  * AS-PATH length
  * Origin
  * MED
  * One of these:
    o Neighboring AS or sub-AS (before the addition of the eiBGP
      Multipath feature)
    o AS-PATH (after the addition of the eiBGP Multipath feature)

In your case neither the Neighboring AS nor the AS-PATH will be the same.
This is the main reason why I never spent time to allow multipath
selection in bgpd. It will only work in very few setups.

-- 
:wq Claudio