Poor methodology

(Re-written from my comment at lwn.net):

This isn't a particularly well-conducted test.

Most notably:

- The article measures power draw under load without measuring performance. A configuration that simply went slower and required more power in total (due to greater time requirements) to
complete a task would thus "win" the test.

- The "power saver" instead of "balanced" profile for 2008 server was used. That's a poor choice for a server under any circumstances and may even use *more* power to complete a given job because it can prevent the CPU from spiking to max performance to finish a job quickly.

- The load test was not likely to make a modern server work hard. Something like a file or database server load generator might have been a better choice. I guess if it was virus-scanning and spam-checking vast piles of mail...

- CPUs, including server CPUs, supported the HLT instruction for reduced CPU power draw well before frequency and voltage stepping arrived on the scene even for laptops. Servers haven't run on full-tilt power max when idle for a *long* time.

- The article falls for the all-too-common pretty but misleading non-zero-y-origin graph trap. Argh!

A more useful test would've defined jobs to do (100 loops of this database load test; process 1 million spam-filtered and virus-scanned emails; etc). It would measure total energy (joules) consumed to complete the task in various power modes. It would then ALSO consider idle draw and full-tilt peak draw - both of which matter for, among other things, cooling needs - as secondary concerns.

Click to read the article this is in response to.