Bill Holler wrote: > > PAD data will be a huge hint to the idle thread for which c-state > it can enter. We have several models to try. :-) Here is what > we are planning: > > CPUs in low power group will always attempts to enter the > deepest c-state possible with a latency some multiple less than > the time till the next cyclic. Idle threads will keep track of > its success rate in entering the deeper c-states. We do not yet > know if success-rate data needs to be used in the low-power- > processor-group c-state policy. > CPUs in the higher power groups have a larger chance a thread > will be dispatched on them before the next cyclic expires. > We are not sure yet how to model this prediction. Success > data will probably be very important here. We might also experiment with tracking the "depth of sleep" for a CPU, and given all else is equal have the dispatcher prefer CPUs in a shallower sleep.
-Eric
