On 01-Jul 14:35, toby huang wrote:
Hi Patrick,
With your new fix patch, you define
perf_constrain_idx = perf_boost_idx,
As my understanding, it means will permit more performance reduction for PC region with the same energy reduction. For example:
boost = 15, energ_delt = -10.
Previous solution: perf_boost_idx = 1, perf_constrain_index = 9,
For PC region, nrg_gain = 5, cap_gain = 1, then capa_delt * 5 - (-10 x 1) > 0 capa_delt > -2
New solution: perf_boost_idx = 1, perf_constrain_index = 1,
For PC region, nrg_gain = 1, cap_gain = 5, then capa_delt * 1 - (-10 x 5) > 0 capa_delt > -50
so, we can see, for new solution, the accepted capacity delta range is more big.
I don't know whether my understanding is correct, if true, it will cause more performance downgrade? am i right?
You are right, but that is the intended design of SchedTune.
If you use low boost values (e.g. 10%) you are more biased towards energy saving, and thus you are willing to accept bigger impact on performances for a certain amount of energy saving.
While, when you boost more (e.g. 80%) than you are more biased toward performance boosting, and thus you are willing to reject big impact on performances for the same amount of energy saving.
Unfortunately, the previous implementation was breaking this assumption.
Cheers Patrick
-- #include <best/regards.h>
Patrick Bellasi