| Message ID | 20230728145808.970594909@infradead.org (mailing list archive) |
|---|---|
| State | RFC, archived |
| Headers | show |
| Series | cpuidle,teo: Improve TEO tick decisions | expand |
On Fri, Jul 28, 2023 at 5:01 PM Peter Zijlstra <peterz@infradead.org> wrote: > > When selecting a state, stop when history tells us 66% of recent idles > were at or below our current state. This is not really about "recent idles" AFAICS. It stops the selection when 66% of the total sum of the "hits" and "intercepts" signals comes from the current state and the states below it. This covers the entire history, while the "recent" ones are counted separately. > Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> > --- > drivers/cpuidle/governors/teo.c | 6 ++++++ > 1 file changed, 6 insertions(+) > > --- a/drivers/cpuidle/governors/teo.c > +++ b/drivers/cpuidle/governors/teo.c > @@ -363,6 +363,7 @@ 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; > @@ -397,6 +398,8 @@ 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 > @@ -427,6 +430,9 @@ 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) > >
On Fri, Jul 28, 2023 at 07:07:02PM +0200, Rafael J. Wysocki wrote: > On Fri, Jul 28, 2023 at 5:01 PM Peter Zijlstra <peterz@infradead.org> wrote: > > > > When selecting a state, stop when history tells us 66% of recent idles > > were at or below our current state. > > This is not really about "recent idles" AFAICS. It stops the > selection when 66% of the total sum of the "hits" and "intercepts" > signals comes from the current state and the states below it. This > covers the entire history, while the "recent" ones are counted > separately. Ah, bad wording perhaps. I found that 'recent' list teo has very hard to use, also the regular state isn't that long lived as per the reply on the previous patch.
--- a/drivers/cpuidle/governors/teo.c +++ b/drivers/cpuidle/governors/teo.c @@ -363,6 +363,7 @@ 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; @@ -397,6 +398,8 @@ 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 @@ -427,6 +430,9 @@ 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)
When selecting a state, stop when history tells us 66% of recent idles were at or below our current state. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> --- drivers/cpuidle/governors/teo.c | 6 ++++++ 1 file changed, 6 insertions(+)