[c-nsp] Per Packet vs Per Destination Load Balancing

[Mark Costlow]

> On Fri, 2004-10-22 at 17:21, Chris Strandt wrote:
> > Is there a performance impact or limitation with per destination load 
> > balancing?
> 
> (not quite answering you question, but giving you another data point) 
> 
> >>From my experience (all of a week:) with balancing two T1 lines (ospf
> +cef).  With per-packet; any traffic can use all the bandwidth.  With
> per-destination; when one link is saturated, new traffic does _not_ tend
> to migrate to the other less used circuit.  On a 2620, I didn't notice
> any significant cpu usage either way.

We ran into a surprising limitation with per-destination load balancing
recently.  Well, not really surprising, except that we didn't think about
it beforehand.

We have some POPs shared with another ISP and we tunnel all of the
traffic back to our data center with GRE tunnels.  There's a router
doing per-destination load balancing of 2 T1s in the path between the two
tunnel endpoints.  The problem is that the tunnel traffic all appears to
be a single flow.  So even though we had many individual customers feeding
into the tunnel (with many different flows), all that traffic ended up
on a single T1, even if the 2nd T1 was empty.

Obvious in hindsight, but a "D'oh!" moment for me.  Switching to per-packet
fixed it, and had very little impact on CPU utilization.

Mark
-- 
Mark Costlow    | Southwest Cyberport | Fax:   +1-505-232-7975
cheeks at swcp.com | Web:   www.swcp.com | Voice: +1-505-232-7992

      "Education is never a waste" - Viscount du Valmont