Cooled by the oft-chilly winds blowing off of Lake Michigan, Chicago was rated in a study conducted last year as the most energy-efficient U.S. city in which to build a datacenter. But the density of Microsoft's datacenter in Northlake is requiring the company to construct three electrical substations that will provide a total of 198 megawatts of electricity for powering and cooling systems, according to a story posted by the Data Center Knowledge online news site.
That's enough electricity to power almost 200,000 homes, and Manos told Data Center Knowledge that about 82 percent of the $500 million bill for the Northlake datacenter is going toward the facility's mechanical and electrical infrastructure.
Long used by the U.S. military, containers filled with preconfigured, ready-to-run servers are being touted as a quicker, more modular way to expand datacenters on the fly than installing racks of servers one by one. Google and Sun have both filed patent claims on server-filled containers, although the former isn't thought to be actively deploying them. Besides Sun, other vendors of container-based setups include IBM, Dell, and Rackable Systems.
Despite its huge size and 24/7 operations, Microsoft's Northlake datacenter won't provide much of a lift to the IT job market in the Chicago area. Manos has said that the new facility would employ only about 30 people, including systems administrators as well as building security and janitorial staffers. In contrast, Google has said that a $600 million datacenter it is building in Council Bluffs, Iowa, will have about 200 employees when it opens next year.
Microsoft's theory, according to a 2007 presentation by Hamilton (download Word document), is that a smaller staff will actually boost the datacenter's reliability. Hamilton claimed that between 20 percent and 50 percent of system outages are due to "human administrative error," and argued that letting malfunctioning hardware die off was a wiser strategy for a redundantly networked datacenter than trying to fix the systems and thus potentially risking a larger failure.
"As parts fail, surviving nodes continue to support the load," Hamilton wrote. "In this modified model, the constituent components are never serviced and the entire module just slowly degrades over time as more and more systems suffer non-recoverable hardware errors." He added that even if 50 of the servers in a 1,000-system module suffered fatal hardware failures, the module would still be "operating with 95 percent of its original design capacity."
Computerworld is an InfoWorld affiliate.