Darshan Purohit wrote:
> I have question about multipath routing. I am running a 2.6.7 kernel 
> (gentoo). 
> I have  a route with three nexthops on the same interface. I see a
> different nexthop being picked for different destination addresses.
> All is fine.

Are you saying that you have three different possible (default) routers / 
gateways as your hosts upstream router / gateway?  If this is the case and you 
are worried about what happens when one of them goes down I have a suggestion 
for you.  If you have control of your upstream routers / gateways  and they are 
running Linux (or possibly something else) you might want to consider 
installing and configuring VRRP (comparable to Cisco's HSRP) to make one 
virtual router that you could configure all the systems on your network to use. 
 VRRP is meant to have two or more routers collectively acting like one virtual 
router.  If the (acting) master router goes down (does not respond to a heart 
beat from the other VRRP router(s)) the next slave router will step up and take 
over the role of the master router.  I think this is done in a close to 
seamless as possible from the client system perspective.  Just a thought.

> Now if one of the nexthop goes down (arp entry times out and arp
> request doesnt get a response), does it remove the nexthop from
> contention and only use the remaining two from this point on ?  What
> about the flows for which this nexthop was chosen in the recent past.
> Would they move to a different nexthop ?

I think that I have read (don't take my word for it though) that the nexthop is 
left in the list of available hops and a different one is chosen.  But like I 
said I'm not sure of this.

> Can anyone please point me to  implementation details about how a
> nexthop is chosen and what is the algorithm used in case of multipath
> routes and how to fine-tune the behaviour.



Grant. . . .
_______________________________________________
LARTC mailing list
[email protected]
http://mailman.ds9a.nl/cgi-bin/mailman/listinfo/lartc

Reply via email to