I've just been testing some changes I made to improve the PAD P-state
transition policy.
Before, PAD would ramp up the max speed as soon as the P-state domain
went non-idle, and then ramp it back down the instant it goes idle again.
Running this script on that kernel, focusing on an idle CPU:
#!/usr/sbin/dtrace -s
#pragma D option quiet
::ess_transition
/curcpu->cpu_id == $1/
{
x++;
}
profile-1s
/x > 0/
{
printf("%d\n", x);
x = 0;
}
I would get ~140 transitions per second. With the new changes, we don't
transition an idle P-state domain up to full speed until the first
thread that begins running there either makes it through one time
quantum, or transitions to another non-idle thread. This gets rid of the
transitions we were doing simply as a result of transient background
kernel activity, reducing the rate down to 2-4 transitions per
second...and 0 per second once the domains are busy....
-Eric
