[c-nsp] Per packet Load balancing

Amol Sapkal amolsapkal at gmail.com
Tue Sep 7 06:45:13 EDT 2004


Hi,


On Mon, 6 Sep 2004 23:14:37 -0400, Rodney Dunn <rodunn at cisco.com> wrote:
> It's better to draw a picture for scenarios like this.
> 
> I think I get what you have done.
> 
> Internet -- R1 ---- PVCa ----\
>                |                         \--CE -- /28
>               R2 ---- PVCb -------/
> 

Yes, this is exactly the case, where R1 connects to Internet and PVCa,
PVCb run from R1, R2 respectively to the CE.


> So to loadbalance from the CE to the internet you
> put two default routes on the CE pointing at PVCa and PVCb.
> 

Yes

> Then to get to the /28 from the internet you put a /28
> route on R1 to PVCa and a second pointing at R2 so R1 has
> equal cost routes (at least from R1's perspective) to get
> to the CE. You put the same /28 route on R2 towards PVCb I
> assume.
> 

Exactly

> Given that, it still will not guarantee you have equal
> load balancing over those links because all it takes is
> one large pair of ip addresses to use all the bandwidth.
> My post a few days ago went over this in detail.
> One flow (src/dst) ip pair could eat up all the bandwidth
> on one of the PVC's.
> 
> Now if your spread of traffic is pretty good between the
> /28 and the Internet you should see *some* load sharing
> but it will not be 50/50 most likely.
> 


Yes, actually on the /28, there is one IP which is a proxy and eats up
almost everything. Even the IP analysis shows that most traffic runs
to Yahoo and Rediffmail.

> Your best bet is to run netflow and see who are the top
> talkers and to verify the path they are taking on R1
> do "sh ip cef exact-route <src> <dst>".
> 
> This is an example where you may have to do per-packet
> if you are required to get more granular on the traffic
> split.
> 

As per my previous mail, the 'ip cef' was initially enabled, but I
disabled it when I saw that particular flows are eating up the
bandwidth on one path.

But, *even after disabling cef*, the load balancing is not occurring.
If I am not wrong, this makes my configuration a 'per-packet' load
balancing.



> Rodney
> 


-- 
Warm Regds,

Amol Sapkal

--------------------------------------------------------------------
An eye for an eye makes the whole world blind 
- Mahatma Gandhi
--------------------------------------------------------------------


More information about the cisco-nsp mailing list