The MPLS WG Archive

[Curtis Villamizar <curtis@avici.com>]

In message <39A1ACFF.50852E4E@marconi.com>, David Charlap writes:
> Erickson Trejo-Reyes wrote:
> >
> > While reading about MPLS specification (all my previous experience is
> > on ATM networks), I have always found statements like: - a path is
> > selected and resources are reserved in every node it traverses... Does
> > that imply some sort of call admission control?
> 
> Any time resources are reserved (on any system, not just MPLS), there
> must be admission control.  Otherwise, a router would be unable to
> actually guarantee the resources it claims to be reserving.
> 
> This is done as a part of the signalling code (either RSVP-TE or
> CR-LDP).
> 
> Of course, when tunnels are signalled for best-effort traffic, then no
> admission control is necessary.

That's one way of doing things and it is the one that is most familiar
to people coming from an ATM background.  btw- Making no reservations
for best-effort traffic and not providing any means for best effort to
be traffic engineeered and route around congestion has been referred
to in the IRTF (see end2end-interest archives) as one component of
"worst-effort" service for best effort traffic.

There are others (or maybe its just me :) that have a different vision
of how to take advantage of what MPLS has to offer.  Please don't
start a flame war over whether there is One True Way to use MPLS.
There are at least two (probably more).  Here is an alternate.

While CAC is fully supported it can essentially be ignored by setting
a very high overbooking factor.  If that is done, some other means to
cause traffic to route around congestion is needed.  Typically the
metric used in the SPF is biased by the ingress to reflect loading
(heavily loaded links are made more costly in the OSPF cost sense or
given an artificially higher metric using ISIS terminology, either way
it is an CR-SPF).  

Behavior is significantly different (from the ATM/CAC way of doing
things) when a major change occurs in the network such as link
failure.  Tunnels will tend to reroute quickly, often more quickly
than information can be flooded back to indicate that bandwidth is
being used up or is no longer available.  Without overbooking, CAC
kicks in and some reservation request fail and must be retried.  With
the high overbooking, these tunnels come up anyway.  If there are
preferred services, they get preferred treatment so the overbooking
doesn't matter to them.  Best effort services temporarily get
squeezed.  The reason for doing things this way is that a very small
amount of loss can cause TCP to back off significantly, therefore
oversubscribing is much more preferrable to having CAC kick in and
losing 100% of the traffic until things settle down.

After the initial layout, if flooding indicates that some of the links
are significantly oversubscribed, then a good adaptivity
implementation can begin to optimize the layout.  Adaptivity is needed
regardless of whether oversubscription is used because some tunnels
which did not traverse a link that went down may need to be moved in
order to achieve a more efficient layout.

There is no One True Way to use MPLS.  This is just another way to use
MPLS for which some interest has been expressed.

Curtis