Efficient datacenter case studies [/ɪˈfɪʃənt ˈdatasenter ˈkeɪs ˈstədiz/] n - Efficiency at hyperscale is no single trick. It is a stack of decisions: where to build, how to cool, how to shape the servers, how to run the software, and how to reckon all of it. The cases below merit reading because each isolates some idea a far smaller site may borrow.
The Cases
- Google’s efficient datacenter design [/ˈgoogle’s ˈeffisient ˈdatasenter ˈdesign/] n - measuring, cooling, and building nearer to renewable power.
- Meta’s efficient datacenter design [/ˈmeta’s ˈeffisient ˈdatasenter ˈdesign/] n - open hardware, free cooling, and a building designed around the workload.
- The Datacenter Beneath the Sea [/ˈmikrosoft’s ˈunderwater ˈdatasenter ˈexperiment/] n - the ocean employed as a heat sink and a controlled atmosphere.
CERN datacenter in Geneva, showing aisle containment and dense rack layout. Photo by Hugovanmeijeren, licensed CC BY-SA 3.0 via Wikimedia Commons.
Lessons That Travel
Certain themes recur, and I record them here:
- Measure first. [/ˈmɪːsʌrɛ ˈfɪrst/] n Nothing improves that is not measured. PUE is a blunt instrument, yet it is a beginning. See Power usage effectiveness.
- Match the site to the climate. [/ˈmætkh ˈthɛ ˈsɪtɛ ˈtɒ ˈthɛ ˈklɪmætɛ/] n Free cooling, evaporative cooling, and immersion all lean upon the local environment. The best design in Finland is not the best design in Singapore.
- Co-design hardware and facility. [/ˈkɒdɛsɪgn ˈhærdwærɛ ˈænd ˈfækɪlɪti/] n A server contrived for a hot aisle is not the server contrived for warm water. The boundary between IT and facilities is an artificial one.
- Automate the response. [/ˈæʌtɒmætɛ ˈthɛ ˈrɛspɒnsɛ/] n The most efficient sites keep no operators chasing alarms. They keep control loops that throttle fans, shift workloads, and summon maintenance before any human has noticed the need.
The custom-device work in these pages is the same idea writ small: measure the specific thing, close the loop where you stand, and learn. See Bespoke Devices for the Datacenter and On Early Detection.
Part of The New Buildout.