Messages in this thread Patch in this message | | | Date | Wed, 26 Jul 2023 22:09:49 +0200 | From | Peter Zijlstra <> | Subject | Re: Stopping the tick on a fully loaded system |
| |
On Wed, Jul 26, 2023 at 08:30:01PM +0200, Rafael J. Wysocki wrote:
> > - The governors teo and menu do the tick_nohz_next_event() check even if > > the CPU is fully loaded and but the check is not for free. > > Let me have a loot at teo in that respect. > > The problem is when tick_nohz_get_sleep_length() should not be called. > The easy case is when the governor would select the shallowest idle > state without taking it into account, but what about the deeper ones? > I guess this depends on the exit latency of the current candidate idle > state, but what exit latency would be low enough? I guess 2 us would > be fine, but what about 10 us, or even 20 us for that matter?
The patch I send here:
https://lkml.kernel.org/r/20230726164958.GV38236@hirez.programming.kicks-ass.net
(which was stuck in a mailqueue :/) tries to address that.
Additionally, I think we can do something like this on top of all that, stop going deeper when 66% of wakeups is at or below the current state.
--- a/drivers/cpuidle/governors/teo.c +++ b/drivers/cpuidle/governors/teo.c @@ -362,6 +362,7 @@ static int teo_select(struct cpuidle_dri unsigned int idx_hit_sum = 0; unsigned int hit_sum = 0; unsigned int tick_sum = 0; + unsigned int thresh_sum = 0; int constraint_idx = 0; int idx0 = 0, idx = -1; bool alt_intercepts, alt_recent; @@ -396,6 +397,8 @@ static int teo_select(struct cpuidle_dri duration_ns = tick_nohz_get_sleep_length(&delta_tick); cpu_data->sleep_length_ns = duration_ns; + thresh_sum = 2 * cpu_data->total / 3; /* 66% */ + /* * Find the deepest idle state whose target residency does not exceed * the current sleep length and the deepest idle state not deeper than @@ -426,6 +429,9 @@ static int teo_select(struct cpuidle_dri if (s->target_residency_ns > duration_ns) break; + if (intercept_sum + hit_sum > thresh_sum) + break; + idx = i; if (s->exit_latency_ns <= latency_req)
| |