Centralized vs. distributed data centers: Green, redundant, both? by Doug Mohney
When it comes to building data centers, organizations are building big and centralized. AOL appears to be bucking the trend with frothing over its Micro Data Center (MDC). Will the need for OpEx efficiency win out over fault tolerance and redundancy?
Clearly, the trend among the big boys is regional centralization. Amazon, Apple, Facebook, Google and Microsoft all build big facilities, typically near the cheapest power they can find. If accessible and relatively affordable, companies opt for green features such as renewable power and environmentally benign cooling -- the latter strategy being to build in a cooler climate and effectively "open the window" to dump or shunt heat out of the facility without having to use as much energy.
Having lots of servers under one roof offers clear economies of scale in terms of management, purchasing and operational expenses. Fewer facilities means less overhead for tracking equipment, power usage and even tax calculations. A pledge to build a big data center also provides leverage with local municipalities with the promise of jobs in exchange for tax breaks.
On the flip side, a large data center can make a significant impact upon water supplies. The Bend Bulletin of Central Oregon reports Apple's expansion in Prineville has lead the company to team up with the city to tap into an underground stream to improve the municipality's water capacity. Apple plans to build two "data halls" covering over 500,000 square feet, so it will need reliable water for cooling. Some facilities need water for making ice to moderate temperature and power costs, creating ice when power is cheaper in the evenings and using it as a literal cold sink during the day to offset air conditioning usage.
Centralized facilities also bring up single-point-of-failure. With increasing near-term weather fluctuations such as Washington D.C.'s summer "derecho" and streaks of hurricanes roaming through Huntsville, Alabama, the impact of a substantial weather or geological event upon Maiden, North Carolina or the aforementioned Prineville, Oregon -- home of data centers for Apple and Facebook -- could cause considerable impact upon the operations and financial bottom line.
If you're a card carrying member of the professionally paranoid, redundancy in data center capacity is necessary in case of a disaster. Virtualization is going to play a key role in large-scale contingency planning. Today, you can move a virtualized process between servers and a bunch of processes off a failing server to another one, either within a data center or to an entirely new facility. In the vaguely defined "future,” being able to relocate data center operations between centers -- a la the "Follow the Sun" principle -- due to a crisis should be standard practice.
I can also see a future where places such as Verne Global's data center in Iceland (yes, Verne is a sponsor of this blog), offer cost-effective near "instant-on" server capacity. A next-generation combination of new servers and software could provide low-power virtualized server backup capacity able to come to life in seconds -- if someone's willing to front some dollars for the initial capital and ongoing operational expenses.
Which brings us back to the AOL "We can put a server anywhere" MDC. I like MDCs for some applications. They would appear to be idea, with a little fine tuning, for applications in the growing markets of Asia and Africa. Designed to work off of minimal and/or no-grid power, a MDC-esque solution offers great promise.
In more developed markets, I'm not so sure if MDCs are the best solution. Certainly, it is good to push out more compute power to the "edge" of the network to avoid bottlenecks. Distributing hundreds to thousands of them to hundreds of different locations sounds like an accounting nightmare of real estate deals -- not to mention a loss of efficiency in green-ness unless you're plugging them all into distributed renewable power.