dmc_ondemand downdifferential

[jimmyhon]

Regarding using a upthreshold of 20 vs 25. Have you considered lowering the downdifferential to get the memory frequency back down to idle?

With the default downdifferential of 20 [1], by changing the upthreshold to 20, governor will stay at the same frequency almost all the time once it's boosted up. [2] (i.e. upthreshold - downdifferential is 0, so the frequency will not drop unless busy_time is 0.)

1. https://gitlab.com/rk3588_linux/rk/kernel/-/blob/linux-5.10-gen-rkr3.6/arch/arm64/boot/dts/rockchip/rk3588s.dtsi#L1289-1290
2. https://gitlab.com/rk3588_linux/rk/kernel/-/blob/linux-5.10-gen-rkr3.6/drivers/devfreq/rockchip_dmc.c#L2810-2815

[jimmyhon]

My closest 7z b -mtt=8 score to dmc governor=performance is using upthreshold/downdifferential=10 and 5 (10_5)

[alarm@cinq ~]$ tail -n2 dmc_*
==> dmc_ondemand_10_5 <==
Avr:             731   2065  15091  |              681   2543  17314
Tot:             706   2304  16202

==> dmc_ondemand_20_10 <==
Avr:             727   2025  14717  |              680   2449  16645
Tot:             703   2237  15681

==> dmc_ondemand_20_20 <==
Avr:             748   1994  14908  |              681   2583  17582
Tot:             714   2289  16245

==> dmc_ondemand_25_20 <==
Avr:             741   1995  14774  |              681   2434  16566
Tot:             711   2214  15670

==> dmc_ondemand_40_20 <==
Avr:             750   1714  12848  |              681   2433  16565
Tot:             715   2074  14706

==> dmc_performance <==
Avr:             724   2015  14571  |              681   2583  17598
Tot:             702   2299  16085

When using upthreshold/downdifferential=20_20, I experience the same issue of the memory clock not dropping back to idle.

When looking at cat /sys/class/devfreq/dmc/trans_stat, the 7z benchmark using (10_5) there around 20 transitions between 1.0 GHz and 2.1 GHz clocks.

[ThomasKaiser]

Sorry for having this ignored for so long. But I wanted to integrate at least one other Rockchip SoC into testing before proceeding (generating settings that provide better performance while not harming idle or 'medium load' consumption too much).

One of my RK3568 machines is available for testing again (BTW: the DMC stuff also affects RK3528) and I have a NetIO powermeter in the lab.

Did you come to different conclusions in the meantime?