In case you aren't suitably impressed by the scale of Amazon Web Services
The video is embedded below (or you can watch it on YouTube here, but here are some facts to whet your appetite:
Networking only represents 8% of monthly AWS operating costs, Hamilton says, but the "cost of networking is escalating relative to the cost of all the other equipment." That is very "anti-Moore," he says.
"At the same time, the ratio of networking to compute is going up," he says. "Partly driven by the fact there is more computer in a given server generationally because Moore is working there, partly because, as the cost of computing falls, the amount of advanced data analytics goes up, which is network intensive."
This pushed AWS to design its own switches, which taught the company a few things. "One, they are a lot cheaper, which isn't a surprise. We were paying $10s of millions on support contracts alone. But surprisingly, availability went up. Why Because suppliers have to aggregate the needs of tons of customers. We don't end up using all the features, and neither do you."
AWS stripped the switch design to the element they needed, and ended up with "an easier problem to solve," which made the devices more reliable.
Although few companies have the resources of these Web-scale guys, this is the basic thinking driving interest in the whitebox/commercial off the shelf (COTS) movements.
Hamilton goes on to outline AWS' basic worldwide architecture:
While most AZs only have two AZs, the AWS US East region has five. And that region uses 82,000+ fiber strands to link the AZs, link the AZs to customers and the Internet.
Amazing stuff. And revisit that growth fact: Every day AWS adds as much new server capacity as Amazon used to support its $7 billion business back in 2004.