Re: Multicast in BGP/MPLS IP VPNs

[Eric Rosen <erosen at cisco.com>]

> Likewise, for 2547 VPNs one need to ask why replication at the
> edge should be ruled out (as it is done by the mVPN draft),

I removed this option because as far as I could tell, no one had ever showed
much interest in it.   It can go back in as an option  if that's what the WG
decides.

> especially, if replication at the edge is sufficient for VPLS. 

I don't think anyone  has ever said that it is sufficient  for VPLS if there
is a large amount of IP multicast traffic; this analysis was never done.

Ingress replication generates "excess  traffic" in proportion to the product
of (a) the amount of received  multicast traffic and (b) the number of sites
to which  that traffic must  be sent, i.e.,  the number of sites  which have
receivers to that traffic. 

"Excess traffic" in  this context would be traffic where  a single link ends
up  carrying more  than one  copy of  each packet,  something  which doesn't
happen if you  use multicast distribution trees.  That's  the whole point of
having  an MDT.   (Though if  one were  to try  to do  all  one's multicasts
through  an RP-rooted shared  tree, without  having the  RP join  the source
tree, one  would be giving up some  of that advantage.)  So  the question is
how much excess traffic gets generated, and what is the impact.

If  the  multicast traffic  becomes  a  significant  fraction of  the  total
traffic, then  this multiplier  effect could  have a real  impact on  the SP
network,  even if  each  VPN or  VPLS instance  has  only a  few sites  with
multicast  receivers.  So  I  have to  suspect  that this  option is  really
limited to situations  in which the amount of  multicast traffic, even after
replication, is not a significant portion of the total traffic.  Essentially
the SP would be telling its customers "you can do  as much multicasting as
you want as long as you don't do too much". 

I think there's some hope that  unknown DA traffic might fall into this "not
too  much" category,  but less  hope that  IP multicast  will fall  into it.

> when an MDT  does not guarantee that the  (multicast) traffic be delivered
> to only  the PEs that  have receivers for  that traffic (like the  case of
> default MDT),  it seems rather  arbitrary to restrict  such MDT to  only a
> single VPN. 

Assigning multiple VPNs to a single  MDT is worth considering, but the trick
of course is to come up with an assignment rule which is not arbitrary.