lies, damn lies, and network traffic statistics

How do traffic engineering and core network design impact the cost and performance of an IP network? I have long seen ad hoc capacity augments as the most blunt and most expensive of tools to improve network performance. But I'm beginning to have doubts, or at least I'm doubting that I can prove my point.

In theory, a traffic-engineered network will be cheaper and perform better than a "native" network. In practice, it seems that demand path and demand volume uncertainty means any IP network backbone has to be significantly overbuilt to achieve the highest level of performance and reliability. If that's the case, then one of the arguments for traffic engineering-- lower network cost-- is invalid.

The problem is that in practice backup paths even in a traffic-engineered network have to be overbuilt. In some failure or surge scenarios, the backup paths are liable to take traffic greater than expected from measurements of normal operating conditions. The solution is the fine art of ad hoc overprovisioning of backbone capacity: something we say we don't do.

Before you think this post is confessional, let me state that I think a traffic-engineered network is less expensive than a "native" network. But I would like to know for sure, and be able to prove what I know. How big do you have to build a "native" network? How big do you have to build the same network using various traffic-engineering schemes? It isn't true in practice that we build it only as big as the traffic-engineering scheme and network statistics say we should. We overbuild it above and beyond what the simulations say we need. How much bigger would the same "native" network be, in practice?

I suspect the answer will show that the costs are similar for different traffic engineering schemes, and the choice of traffic engineering scheme depends mostly on operations, vendor selection, and product specifications. In other words, there is probably little reason to change traffic engineering schemes once one is in place, because the cost and performance of various schemes is almost equal. Presuming that any reasonable choice can be made to perform effectively, given enough bandwidth, consider that the IP backbone equipment in a network is only about 10% of the total network cost. The most expensive part of the network is at the edge, where the customer services are provisioned. Even doubling the cost of the IP core would only add about 10% to the total network cost. When you're trying to live within your budget, that's a big deal. But when you're talking about the competitive advantage of one scheme over another one that's peanuts, and the winning carrier will have better reach, customer service, and performance, not a better traffic engineering scheme.

But I can't prove it, and that's the point of this post.