Update on DeviceKit

[Matthew Garrett]

On Thu, May 08, 2008 at 03:40:07PM +0200, Holger Macht wrote:
> On Thu 08. May - 14:27:29, Matthew Garrett wrote:
> > Well, yes, but what does that mean? You can't rate performance on a 
> > scale of 0-100 - there's multiple factors at play. Is performance a 
> > latency issue? A raw power one? What should the thresholds for ondemand 
> > be? If you're altering multiple factors, then for some workloads a 
> > higher value may yield lower performance. If you're only altering one 
> > factor, you may not be obtaining maximum performance for that power 
> > consumption. Trying to squash this into a linear scale doesn't work.
> 
> You always get more performance for the price of more power
> consumption. And having less performance results in lower power
> consumption and in a non ideal world, you might cope with more latency. At
> least this should be the goal. How can you say it doesn't work? It _does_.

Here are two cases:

1) Limit the maximum frequency to 1.4GHz, minimum is 800MHz. The 
ondemand sampling rate is 0.1 seconds. A job starts, requiring 
0.5GHz/seconds of work. After 0.1 seconds, the cpu jumps to 1.4GHz. The 
job is done in 0.3 seconds and the cpu returns to idle.

2) Don't limit the maximum frequency. The minimum frequency is limited 
to 1GHz (higher performance). ondemand is sampling at an interval of 1 
second. The same job is started. It finishes in 0.5 seconds, and 
ondemand never scales the processor because the load has vanished by the 
time it samples. The cpu then returns to idle.

(2) consumes more power, but provides lower performance. Yes, I'm not 
comparing identical cases, but that's the point. There's more than one 
factor involved, and you have no way of telling which will result in 
better power consumption and performance without analysing the 
workloads.
-- 
Matthew Garrett | mjg59 at srcf.ucam.org