On 7 February 2011 12:55, Vishwanath Sripathy vishwanath.sripathy@linaro.org wrote:

Vincent,

On Mon, Feb 7, 2011 at 3:55 PM, Vincent Guittot vincent.guittot@linaro.org wrote:

...

Hi vishwa,

I have passed cpufreq-bench on my platform. The results below have been reached with a sampling rate of 20ms and a sampling_down factor set to 10.

I have a question about the sampling_down feature : The ondemand delay value is calculated before calling dbs_check_cpu, but dbs_check_cpu can modify rate_mult. This implies that if the rate_mult is set to 1 in dbs_check_cpu because we set something else than the max frequency, the next sampling period will be sampling_rate*sampling_down_factor (the one used for max freq) but the frequency can be the lowest one. We should rather calculate the delay after dbs_check_cpu or i miss something ?

Good point. Yes, with current implementation, change in rate_mult will have effect only after next do_dbs_timer is called and the point raised by you is perfectly valid. Why don't you post this query to cpu_freq mailing list?

ok, i'm going to send it

Vishwa

...

Vincent

Test results with current implementation, 20ms sampling rate, sampling_down = 10: _round 1: 59.14% _round 2: 56.87% _round 3: 77.31% _round 4: 88.44% _round 5: 89.56% _round 6: 83.17% _round 7: 89.87% _round 8: 94.33% _round 9: 92.98% _round 10: 91.12%

Test results when the delay is calculated after dbs_check_cpu, 20ms sampling rate, sampling_down = 10  : _round 1: 86.75% _round 2: 91.18% _round 3: 93.23% _round 4: 96.70% _round 5: 95.87% _round 6: 96.04% _round 7: 97.39% _round 8: 97.37% _round 9: 98.88% _round 10: 97.78%