Frequency, heat, and idleness
Frequency, heat, and idleness
Posted Apr 6, 2020 4:12 UTC (Mon) by florianfainelli (subscriber, #61952)In reply to: Frequency, heat, and idleness by epa
Parent article: Frequency-invariant utilization tracking for x86
> On the other hand if it's almost idle but has a few occasional bursts of activity, you may want to get those bursts done with the highest possible clock frequency.
In premise yes, but if you factor in voltage scaling to support that high frequency the time to move from idle to a higher frequency + higher voltage may outweigh choosing a slightly lower frequency + voltage point. You need to ensure you have converged to a higher voltage before you can raise the frequency otherwise, well bad things could happen. A lot of this is clearly magic on the hardware side, or rather the magician spent a lot of time in the lab to come up with something adequate. The key for the OS to make good decisions is almost always the same though: you need a cost function that is as accurate as possible.