North American Network Operators Group

Date Prev | Date Next | Date Index | Thread Index | Author Index | Historical

Re: More Sidgemore on per-bit pricing (fwd)

  • From: Vadim Antonov
  • Date: Tue Dec 08 18:44:41 1998

Eric Dean wrote:
> > Please tell me what happens when a link carrying a million MPLS-ed
> > flows does flap.
> The same thing that happens in an SVC environment.  Routing reconverges
> and the pain is felt by all.

Yep, and how much time does it take?  (A: with simplistic oblivious VC
routing which was shown to produce absymally bad utilization in general
topologies it'll take O(N), optimal VC routing is NP-complete (i.e. at
least O(exp(N)); there's a number of heuristics working in polinomial
time).  N here stands for number of end-points in the network.

For comparison, rerouting in per-hop routing network takes O(log N).

Yeah, MPLS may work in private networks; but assuming it'll be of any
use in public Internet backbones is a serious delusion.  The only way
to make it work is to avoid dynamic rerouting by building a double-redundant
full mesh-like network (by mapping PVCs in SONET, for example). 

That is exactly what telephone companies are doing for voice SVCs. 
Unfortunately it also wastes at least 70% of useable bandwidth.

In other words, use SVCs of any kind - and if you want to build a large 
reliable network, you're up to sacrificing most of your network's capacity.
_Not_ doing SVCs provides enough "spare" capacity to provide much better
quality of service to all customers.

(There's one mode of usage in MPLS, downstream demand tag allocation, which
is equivalent to destination-only forwarding; all it is doing is eliminating
IP address lookup.  Which can easily be done in 1.5 memory reads on average;
so the "savings" here are not worth the additional complexity (and related
bugs) of a MPLS code.  Another usage of MPLS in large networks is to
create arcane routing policies (for example, to circumvent "routing fish"),
but, frankly, i never had any problem of that kind which couldn't be fixed
with a length of wire and a crimp tool.)

MPLS shares one common property with ATM: it introduces a lot of complexity
which can be gainfully used only in places where traffic is not very large.
The complexity _always_ has a way of biting back, so the net result is
networks which are less manageable, software which has more bugs, 
interesting failure modes and guaranteed employment for network consultants.

Which is all what that MPLS hype is about. This one will come to pass, too.

I guess many folks fogrot the Magic Rule Of Reliable Networking:

                       Keep It Simple, Stupid.