[s_kanev]

Yeah, cpu idleness is pretty fascinating and, as it turns out, quite important. Shameless plug, I just published a paper [1], showing that if you mess up how deeply you go to sleep in cpuidle, you can loose up to ~15% latency in datacenter services like websearch.

[1] http://www.skanev.org/papers/iiswc14ep.pdf

[amluto]

This reminds me: if you're stuck on a Linux kernel before 3.9 and you're using Sandy Bridge hardware or newer, turn off C1E in BIOS. The "C1E" setting controls a misfeature in which the CPU will look at the OS's request for a light sleep and say "Ha! You must have asked for light sleep by accident! I'm going to go all the way to sleep instead!" This causes surprisingly severe performance problems (we've seen milliseconds of unexplained random hiccups).

Linux 3.9 and higher is immune: it just forcibly disables this feature.

[makomk]

Hah. I guess that's a major benchmark win - decreases power usage in common cases, and few CPU benchmarks test wake-from-sleep latency either directly or indirectly.