Dave Cooper <dcooper@eli.net> wrote:

The problem is not capacity planning. Most large backbones have dedicated resources to determine 6-12 month needs in their cores. There are good tools to help you do that, not to mention good ole' regression analysis along with programs like Wandl. But those statistical tools will only go for 12 months out with a "decent" deviation. Once you go any further, you are grabbing at straws.

The rule of thumb: build as big network as you can and then connect only as many customers as it can sustain :) Seriously, TCP/IP can accomodate very gross mistakes in capacity planning - up to an order of magnitude. Trivial trend analysis and random traffic distribution gives predictions which are good enough. In fact, there are no good ways to measure demand for data traffic -- the actual traffic measurements reflect the current available capacity and effects of routing decisions more than anything else, and only show pathologically bad capacity problems.

The problem lies with the provisioning times for circuits at the OC-3/OC-12/OC-48 level.

That problem should have non-technical solution. Fix telco bureaucracy. Or just let the sluggish ones die; everybody will be better off for that.

This creates the need to use existing resources to reduce latency and delay until you can bring the physical topology to specification.

This creates the need to worry about more than the short term revenue, and to actually listen to what engineers have to say. The problem is not technological, and no technological fix is going to cure it - any such fix only masks the symptoms, simply prolonging the agony and draining scarce engineering resources.

I highly doubt the big players are looking at MPLS to bail themselves out of the physical topology constraints (although that might be a side benefit for both the provider AND the customer).

They're introducing a whole new can of worms complexity-wise for something which at best promises 10-20% improvement (bandwidth utilization-wise; the latency-related actual performance degradation makes even that benefit highly doubtful). 2 months of advantage just do not worth it. Customers increasingly are limited by server capacity, and worry a lot more about availability than about backbone capacity. --vadim