The MPLS WG Archive

[ravikant@NASBPD01BS.ntc.nokia.com (Ravikanth Rayadurgam NRC/Bos)]

Grenville,

Thanks for your reply. I am aware of the improved performance as you have explained it, and at that qualitative level I do see where the improved performance comes from. I guess I was looking for a little more quantitative answer to my question.

As I had indicated in my original post, I was just curious to know how much this improved performance impacts network performance etc., say from an end-to-end perspective. I still feel that it would be helpful to understand this issue.

Comments?

Regards,

 -Ravikanth


 ----------
From: gja
To: Ravikanth Rayadurgam (NRC/Bos)
Cc: mpls; Ravikanth
Subject: Re: Performance question
Date: Wednesday, July 30, 1997 5:22PM



Ravikanth,
 [..]
> I have a question regarding the MPLS work. In Section 1.2 of the
> framework document (draft-ietf-mpls-framework-00.txt), one of the
> requirements is to "improve forwarding performance." I am a little
> unclear on the exact improvements that we are looking for/hope to
> obtain. This may be a point worth discussing.

Here's how it works #1:

 - Take a router whose forwarding table no longer fits in
   the cpu's fastest cache.
 - Observe that cache-misses are causing throughput to fall
   through the floor.
 - Observe that if the forwarding cache was populated with
   labels, you could fit all the labels ('cos they are smaller)
   in the fastest cache. No more cache-misses.
 - Observe the 'improved forwarding performance' as the
   router gets back up off the floor and goes about as
   fast as it used to.

Here's how it works #2:

 - Take a router whose forwarding engine simply isn't fast.
   Throw away the forwarding engine.
 - Observe that if you could label your traffic before it
   reaches the (now gutted) router, you could pass it through
   an ATM switch fabric (we have plenty of these lying around,
   the industry has been building them for awhile now). The
   ATM switch fabric moves data through itself pretty fast.
 - Observe the 'improved forwarding performance'.

The rest of MPLS is about how we ensure labels are assigned
in a meaningful manner across any given cloud of LSRs.

Oh, in case you're wondering, the initial labeling of
IP packets must be done by full routers. You need to
ensure your MPLS cloud architecture is such that the edge
LSRs dont appear at points where traffic load is (a) higher
than they can handle as normal routers, or (b) the forwarding
table is too big for their cache.

(In the case of ATM based MPLS, the labeling is really "fragment,
and label the fragments", and the fragments are called cells
by some, but that's a minor syntactic point.)

cheers,
gja
________________________________________________________________________
Grenville Armitage                        Bell Labs, Lucent Technologies
http://nj5.injersey.com/~gja