The MPLS WG Archive

[Curtis Villamizar <curtis@workhorse.fictitious.org>]

In message <3A54E972.9C7746D6@nortelnetworks.com>, "Cheng-Yin Lee" writes:
> Curtis,
> 
> Curtis Villamizar wrote:
> 
> > There were a number of motivations for getting rid of ATM in IP
> > networks.  One was the cell tax which at a typical 20% would not alone
> > kill ATM.  Another was SAR.  You can't get fast ATM router interfaces
> > because SAR speed and SAR buffering becomes a problem.  This alone was
> > enough to kill ATM.
> >
> > Perhaps even larger was the problems of independent control planes and
> > the effect on IGP scaling.  Flooding is needed for reliability in the
> > IGP and flooding in a full mesh of N routers has some N^3 properties
> > with regard to messages sent when a router in the mesh fails and N^2
> > when a VC fails.  Attempts to constrain the flooding tended to slow
> > convergence and sometimes cause long lived IGP inconsistency when
> > change occurred.  This is not a problem is the flooding follows the
> > physical interconnections as it does with MPLS since the number of
> > adjacencies per node drops dramatically.
> 
> Over time , would there be any inclinations to add direct connectivity betwee
> n B
> and H, and then over F and D, and so on, leading to perhaps densely meshed
> networks in the long run ? What is your opinion on this?
> 
>                           B-------C-------D
>                         /                                          \
>                       /                                               \
>                  A                                                    E
>                     \                                                /
>                       \                                            /
>                        F--------G--------H
> 
> One could use a common control plane and may still run into IGP scaling issue
> s
> because the network is densely meshed, right?


Densely meshed in a 1,000 node network might mean 10-15 adjacencies
per node.  That is still a lot less than full mesh of 999 adjacencies
per node if IP over ATM were used.


> > The IGP scaling problem was address by ISPs by partitioning their
> > network into core and regions but this reduced the effectivenes of TE.
> 
> Could you kindly elaborate your thoughts on this?


Typically a provider would partition the topology into N regions where
N is 10s, then run VCs within an ATM core that interconnected the
regions and VCs within each region.  The routers had no visibility
into TE state so if there was mutliple routers bordering the core and
a given region and one of those core routers had a much less congested
path to a destination (recognizing that congestion was more likely
within specific regions than with the core) then the core routers had
no good way to pick the right edge router and get on the VC with a
cleaner shot at the destination.  [Its late.  I hope this is clear.]


> thanks,
> cyl

Curtis