cpuidle: menu: Add residency threshold for non-polling state selection
PATCH v3 1/1 cpuidle: menu: Add residency threshold for non-polling state selection
On virtualized PowerPC (pseries) systems, where only one polling state (Snooze) and one deep state (CEDE) are available, selecting CEDE when the predicted idle duration exceeds the target residency of the CEDE state can hurt performance. In such cases, the entry/exit overhead of CEDE outweighs the power savings, leading to unnecessary state transitions and higher latency.
Menu governor currently contains a special-case rule that prioritizes the first non-polling state over polling, even when its target residency is much longer than the predicted idle duration. On PowerPC/pseries, where the gap between the polling state (Snooze) and the first non-polling state (CEDE) is large, this behavior causes performance regressions.
This patch refines the special case by adding an extra requirement:
the first non-polling state may only be chosen if its target_residency_ns is below the defined RESIDENCY_THRESHOLD_NS. If this condition is not met, the non-polling state is not selected, and polling state is retained instead.
This change is limited to the single special-case condition for the first non-polling state. The general state selection logic in the menu governor remains unchanged.
Performance improvement observed with pgbench on PowerPC (pseries) system:
+---------------------------+------------+------------+------------+ | Metric | Baseline | Patched | Change (%) | +---------------------------+------------+------------+------------+ | Transactions/sec (TPS) | 495,210 | 536,982 | +8.45% | | Avg latency (ms) | 0.163 | 0.150 | -7.98% | +---------------------------+------------+------------+------------+ CPUIdle state usage: +--------------+--------------+-------------+ | Metric | Baseline | Patched | +--------------+--------------+-------------+ | Total usage | 12,735,820 | 13,918,442 | | Above usage | 11,401,520 | 1,598,210 | | Below usage | 20,145 | 702,395 | +--------------+--------------+-------------+
Above/Total and Below/Total usage percentages which indicates mispredictions:
+------------------------+-----------+---------+ | Metric | Baseline | Patched | +------------------------+-----------+---------+ | Above % (Above/Total) | 89.56% | 11.49% | | Below % (Below/Total) | 0.16% | 5.05% | | Total cpuidle miss (%) | 89.72% | 16.54% | +------------------------+-----------+---------+
The results show that restricting non-polling state selection to cases where its residency is within the threshold reduces mispredictions, lowers unnecessary state transitions, and improves overall throughput.
---
v2: https://lore.kernel.org/all/[email protected]/
Changes in v2 -> v3:
- Modifed the patch following Rafael's feedback, incorporated a residency threshold check
(s->target_residency_ns < RESIDENCY_THRESHOLD_NS) as suggested. - Updated commit message accordingly. --- drivers/cpuidle/governors/menu.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/cpuidle/governors/menu.c +5 -3
@@ -316,11 +316,13 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,if (s->target_residency_ns > predicted_ns) {/* -* Use a physical idle state, not busy polling, unless -* a timer is going to trigger soon enough. +* Use a physical idle state instead of busy polling +* if the next timer doesn't expire soon and its +* target residency is below the residency threshold. */ if ((drv->states[idx].flags &CPUIDLE_FLAG_POLLING ) && - s->target_residency_ns <= data->next_timer_ns) { + s->target_residency_ns <= data->next_timer_ns && + s->target_residency_ns <RESIDENCY_THRESHOLD_NS ) { predicted_ns = s->target_residency_ns; idx = i;break ;
Signature · 1 line
2.50.1
Quoted 78 lines · Aboorva Devarajan
On virtualized PowerPC (pseries) systems, where only one polling state (Snooze) and one deep state (CEDE) are available, selecting CEDE when the predicted idle duration exceeds the target residency of the CEDE state can hurt performance. In such cases, the entry/exit overhead of CEDE outweighs the power savings, leading to unnecessary state transitions and higher latency.
Menu governor currently contains a special-case rule that prioritizes the first non-polling state over polling, even when its target residency is much longer than the predicted idle duration. On PowerPC/pseries, where the gap between the polling state (Snooze) and the first non-polling state (CEDE) is large, this behavior causes performance regressions.
This patch refines the special case by adding an extra requirement:
the first non-polling state may only be chosen if itstarget_residency_nsis below the definedRESIDENCY_THRESHOLD_NS. If this condition is not met, the non-polling state is not selected, and polling state is retained instead.This change is limited to the single special-case condition for the first non-polling state. The general state selection logic in the menu governor remains unchanged.
Performance improvement observed with pgbench on PowerPC (pseries) system:
+---------------------------+------------+------------+------------+ | Metric | Baseline | Patched | Change (%) | +---------------------------+------------+------------+------------+ | Transactions/sec (TPS) | 495,210 | 536,982 | +8.45% | | Avg latency (ms) | 0.163 | 0.150 | -7.98% | +---------------------------+------------+------------+------------+ CPUIdle state usage: +--------------+--------------+-------------+ | Metric | Baseline | Patched | +--------------+--------------+-------------+ | Total usage | 12,735,820 | 13,918,442 | | Above usage | 11,401,520 | 1,598,210 | | Below usage | 20,145 | 702,395 | +--------------+--------------+-------------+Above/Total and Below/Total usage percentages which indicates mispredictions:
+------------------------+-----------+---------+ | Metric | Baseline | Patched | +------------------------+-----------+---------+ | Above % (Above/Total) | 89.56% | 11.49% | | Below % (Below/Total) | 0.16% | 5.05% | | Total cpuidle miss (%) | 89.72% | 16.54% | +------------------------+-----------+---------+The results show that restricting non-polling state selection to cases where its residency is within the threshold reduces mispredictions, lowers unnecessary state transitions, and improves overall throughput.
Signed-off-by Aboorva Devarajan <[email protected]>---
v2: https://lore.kernel.org/all/[email protected]/
Changes in v2 -> v3:
- Modifed the patch following Rafael's feedback, incorporated a residency threshold check(s->target_residency_ns < RESIDENCY_THRESHOLD_NS) as suggested. - Updated commit message accordingly. --- drivers/cpuidle/governors/menu.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/cpuidle/governors/menu.c +1
-2@@ -316,11 +316,13 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,if (s->target_residency_ns > predicted_ns) {/* -* Use a physical idle state, not busy polling, unless -* a timer is going to trigger soon enough. +* Use a physical idle state instead of busy polling
drivers/cpuidle/governors/menu.c static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, · line 328
Quoted 10 lines
drivers/cpuidle/governors/menu.c +4
-1+* if the next timer doesn't expire soon and its +* target residency is below the residency threshold. */ if ((drv->states[idx].flags &CPUIDLE_FLAG_POLLING ) && - s->target_residency_ns <= data->next_timer_ns) { + s->target_residency_ns <= data->next_timer_ns && + s->target_residency_ns <RESIDENCY_THRESHOLD_NS ) { predicted_ns = s->target_residency_ns; idx = i;break ;
To me that seems the least intrusive way the issue for your platform. Rafael, can you live with this?
FWIW
Rafael J. Wysocki
maintainerQuoted 85 lines · Christian Loehle
Quoted 83 lines · Aboorva Devarajan
On virtualized PowerPC (pseries) systems, where only one polling state (Snooze) and one deep state (CEDE) are available, selecting CEDE when the predicted idle duration exceeds the target residency of the CEDE state can hurt performance. In such cases, the entry/exit overhead of CEDE outweighs the power savings, leading to unnecessary state transitions and higher latency.
Menu governor currently contains a special-case rule that prioritizes the first non-polling state over polling, even when its target residency is much longer than the predicted idle duration. On PowerPC/pseries, where the gap between the polling state (Snooze) and the first non-polling state (CEDE) is large, this behavior causes performance regressions.
This patch refines the special case by adding an extra requirement:
the first non-polling state may only be chosen if itstarget_residency_nsis below the definedRESIDENCY_THRESHOLD_NS. If this condition is not met, the non-polling state is not selected, and polling state is retained instead.This change is limited to the single special-case condition for the first non-polling state. The general state selection logic in the menu governor remains unchanged.
Performance improvement observed with pgbench on PowerPC (pseries) system:
+---------------------------+------------+------------+------------+ | Metric | Baseline | Patched | Change (%) | +---------------------------+------------+------------+------------+ | Transactions/sec (TPS) | 495,210 | 536,982 | +8.45% | | Avg latency (ms) | 0.163 | 0.150 | -7.98% | +---------------------------+------------+------------+------------+ CPUIdle state usage: +--------------+--------------+-------------+ | Metric | Baseline | Patched | +--------------+--------------+-------------+ | Total usage | 12,735,820 | 13,918,442 | | Above usage | 11,401,520 | 1,598,210 | | Below usage | 20,145 | 702,395 | +--------------+--------------+-------------+Above/Total and Below/Total usage percentages which indicates mispredictions:
+------------------------+-----------+---------+ | Metric | Baseline | Patched | +------------------------+-----------+---------+ | Above % (Above/Total) | 89.56% | 11.49% | | Below % (Below/Total) | 0.16% | 5.05% | | Total cpuidle miss (%) | 89.72% | 16.54% | +------------------------+-----------+---------+The results show that restricting non-polling state selection to cases where its residency is within the threshold reduces mispredictions, lowers unnecessary state transitions, and improves overall throughput.
Signed-off-by Aboorva Devarajan <[email protected]>---
v2: https://lore.kernel.org/all/[email protected]/
Changes in v2 -> v3:
- Modifed the patch following Rafael's feedback, incorporated a residency threshold check(s->target_residency_ns < RESIDENCY_THRESHOLD_NS) as suggested. - Updated commit message accordingly. --- drivers/cpuidle/governors/menu.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/cpuidle/governors/menu.c +3
-3@@ -316,11 +316,13 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,if (s->target_residency_ns > predicted_ns) {/* -* Use a physical idle state, not busy polling, unless -* a timer is going to trigger soon enough. +* Use a physical idle state instead of busy polling +* if the next timer doesn't expire soon and its +* target residency is below the residency threshold. */ if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) && - s->target_residency_ns <= data->next_timer_ns) {
Quoted 8 lines
drivers/cpuidle/governors/menu.c static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, · line 328
Quoted 5 lines
drivers/cpuidle/governors/menu.c +2
-0+ s->target_residency_ns <= data->next_timer_ns && + s->target_residency_ns < RESIDENCY_THRESHOLD_NS) { predicted_ns = s->target_residency_ns; idx = i;break ;To me that seems the least intrusive way the issue for your platform. Rafael, can you live with this?
Yes, it's fine.
Quoted 1 line
Reviewed-by Christian Loehle <[email protected]>
Rafael J. Wysocki
maintainerPlease change the subject of the patch to something like "cpuidle:
menu: Use residency threshold in polling state override decisions" which more precisely reflects the patch purpose IMV.
Quoted 4 lines · Aboorva Devarajan
On virtualized PowerPC (pseries) systems, where only one polling state (Snooze) and one deep state (CEDE) are available, selecting CEDE when the predicted idle duration exceeds the target residency of the CEDE
If the target residency is exceeded by the predicted idle duration, it should be fine to select the given state.
Did you really mean "less than" here? That would be consistent with the code change.
Quoted 69 lines
state can hurt performance. In such cases, the entry/exit overhead of CEDE outweighs the power savings, leading to unnecessary state transitions and higher latency.
Menu governor currently contains a special-case rule that prioritizes the first non-polling state over polling, even when its target residency is much longer than the predicted idle duration. On PowerPC/pseries, where the gap between the polling state (Snooze) and the first non-polling state (CEDE) is large, this behavior causes performance regressions.
This patch refines the special case by adding an extra requirement:
the first non-polling state may only be chosen if itstarget_residency_nsis below the definedRESIDENCY_THRESHOLD_NS. If this condition is not met, the non-polling state is not selected, and polling state is retained instead.This change is limited to the single special-case condition for the first non-polling state. The general state selection logic in the menu governor remains unchanged.
Performance improvement observed with pgbench on PowerPC (pseries) system:
+---------------------------+------------+------------+------------+ | Metric | Baseline | Patched | Change (%) | +---------------------------+------------+------------+------------+ | Transactions/sec (TPS) | 495,210 | 536,982 | +8.45% | | Avg latency (ms) | 0.163 | 0.150 | -7.98% | +---------------------------+------------+------------+------------+ CPUIdle state usage: +--------------+--------------+-------------+ | Metric | Baseline | Patched | +--------------+--------------+-------------+ | Total usage | 12,735,820 | 13,918,442 | | Above usage | 11,401,520 | 1,598,210 | | Below usage | 20,145 | 702,395 | +--------------+--------------+-------------+Above/Total and Below/Total usage percentages which indicates mispredictions:
+------------------------+-----------+---------+ | Metric | Baseline | Patched | +------------------------+-----------+---------+ | Above % (Above/Total) | 89.56% | 11.49% | | Below % (Below/Total) | 0.16% | 5.05% | | Total cpuidle miss (%) | 89.72% | 16.54% | +------------------------+-----------+---------+The results show that restricting non-polling state selection to cases where its residency is within the threshold reduces mispredictions, lowers unnecessary state transitions, and improves overall throughput.
Signed-off-by Aboorva Devarajan <[email protected]>---
v2: https://lore.kernel.org/all/[email protected]/
Changes in v2 -> v3:
- Modifed the patch following Rafael's feedback, incorporated a residency threshold check(s->target_residency_ns < RESIDENCY_THRESHOLD_NS) as suggested. - Updated commit message accordingly. --- drivers/cpuidle/governors/menu.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/cpuidle/governors/menu.c +0
-0
drivers/cpuidle/governors/menu.c static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, · line 321
Quoted 8 lines
drivers/cpuidle/governors/menu.c +3
-2@@ -316,11 +316,13 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev,if (s->target_residency_ns > predicted_ns) {/* -* Use a physical idle state, not busy polling, unless -* a timer is going to trigger soon enough. +* Use a physical idle state instead of busy polling +* if the next timer doesn't expire soon and its +* target residency is below the residency threshold.
I would rephrase this somewhat, like this:
* Use a physical idle state instead of busy polling so long as
* its target residency is below the residency threshold and the
* next timer doesn't expire soon.
drivers/cpuidle/governors/menu.c static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, · line 325
Quoted 5 lines
drivers/cpuidle/governors/menu.c +2
-1*/ if ((drv->states[idx].flags &CPUIDLE_FLAG_POLLING ) && - s->target_residency_ns <= data->next_timer_ns) { + s->target_residency_ns <= data->next_timer_ns && + s->target_residency_ns <RESIDENCY_THRESHOLD_NS ) {
And maybe adjust the checks ordering here.
The point is that on the example platform in question s->target_residency_ns is always above RESIDENCY_THRESHOLD_NS, so it is never really necessary to check data->next_timer_ns in which case the HW should be able to optimize this.
Quoted 4 lines
drivers/cpuidle/governors/menu.c +0
-1predicted_ns = s->target_residency_ns; idx = i;break ; --
On Wed, 2025-09-10 at 12:47 +0200, Rafael J. Wysocki wrote:
Hi Rafael,
Quoted 3 lines
Please change the subject of the patch to something like "cpuidle:
menu: Use residency threshold in polling state override decisions" which more precisely reflects the patch purpose IMV.
Sure, I will change this.
Quoted 7 lines · Aboorva Devarajan
Quoted 4 lines
On virtualized PowerPC (pseries) systems, where only one polling state (Snooze) and one deep state (CEDE) are available, selecting CEDE when the predicted idle duration exceeds the target residency of the CEDE
Quoted 6 lines
If the target residency is exceeded by the predicted idle duration, it should be fine to select the given state.
Did you really mean "less than" here? That would be consistent with the code change.
yes, I meant "less than" here, will change it.
Quoted 80 lines
Quoted 69 lines
state can hurt performance. In such cases, the entry/exit overhead of CEDE outweighs the power savings, leading to unnecessary state transitions and higher latency.
Menu governor currently contains a special-case rule that prioritizes the first non-polling state over polling, even when its target residency is much longer than the predicted idle duration. On PowerPC/pseries, where the gap between the polling state (Snooze) and the first non-polling state (CEDE) is large, this behavior causes performance regressions.
This patch refines the special case by adding an extra requirement:
the first non-polling state may only be chosen if itstarget_residency_nsis below the definedRESIDENCY_THRESHOLD_NS. If this condition is not met, the non-polling state is not selected, and polling state is retained instead.This change is limited to the single special-case condition for the first non-polling state. The general state selection logic in the menu governor remains unchanged.
Performance improvement observed with pgbench on PowerPC (pseries) system:
+---------------------------+------------+------------+------------+ > Metric | Baseline | Patched | Change (%) | +---------------------------+------------+------------+------------+ > Transactions/sec (TPS) | 495,210 | 536,982 | +8.45% | > Avg latency (ms) | 0.163 | 0.150 | -7.98% | +---------------------------+------------+------------+------------+ CPUIdle state usage: +--------------+--------------+-------------+ > Metric | Baseline | Patched | +--------------+--------------+-------------+ > Total usage | 12,735,820 | 13,918,442 | > Above usage | 11,401,520 | 1,598,210 | > Below usage | 20,145 | 702,395 | +--------------+--------------+-------------+Above/Total and Below/Total usage percentages which indicates mispredictions:
+------------------------+-----------+---------+ > Metric | Baseline | Patched | +------------------------+-----------+---------+ > Above % (Above/Total) | 89.56% | 11.49% | > Below % (Below/Total) | 0.16% | 5.05% | > Total cpuidle miss (%) | 89.72% | 16.54% | +------------------------+-----------+---------+The results show that restricting non-polling state selection to cases where its residency is within the threshold reduces mispredictions, lowers unnecessary state transitions, and improves overall throughput.
Signed-off-by Aboorva Devarajan <[email protected]>---
v2: https://lore.kernel.org/all/[email protected]/
Changes in v2 -> v3:
- Modifed the patch following Rafael's feedback, incorporated a residency threshold check (s->target_residency_ns<RESIDENCY_THRESHOLD_NS) as suggested.
- Updated commit message accordingly.
---
drivers/cpuidle/governors/menu.c | 8 +++++---1 file changed, 5 insertions(+), 3 deletions(-)drivers/cpuidle/governors/menu.c +0
-0drivers/cpuidle/governors/menu.c static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, · line 321
Quoted 8 lines
drivers/cpuidle/governors/menu.c +3
-2@@ -316,11 +316,13 @@ static int menu_select(struct cpuidle_driver *drv, struct cpuidle_device *dev, if (s->target_residency_ns > predicted_ns) { /* -* Use a physical idle state, not busy polling, unless -* a timer is going to trigger soon enough. +* Use a physical idle state instead of busy polling +* if the next timer doesn't expire soon and its +* target residency is below the residency threshold. I would rephrase this somewhat, like this:
Quoted 3 lines
* Use a physical idle state instead of busy polling so long as
* its target residency is below the residency threshold and the
* next timer doesn't expire soon.
Sure, will change this.
Quoted 7 lines
Quoted 6 lines
drivers/cpuidle/governors/menu.c +0
-0*/ if ((drv->states[idx].flags &
CPUIDLE_FLAG_POLLING) && -s->target_residency_ns<=data->next_timer_ns) { +s->target_residency_ns<=data->next_timer_ns&& +s->target_residency_ns<RESIDENCY_THRESHOLD_NS) {
Quoted 6 lines
And maybe adjust the checks ordering here.
The point is that on the example platform in question
s->target_residency_nsis always aboveRESIDENCY_THRESHOLD_NS, so it is never really necessary to checkdata->next_timer_nsin which case the HW should be able to optimize this.
That's right, I will change the condition as follows:
if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
s->target_residency_ns < RESIDENCY_THRESHOLD_NS &&
s->target_residency_ns <= data->next_timer_ns) { Quoted 5 lines
Quoted 5 lines
predicted_ns=s->target_residency_ns; idx = i;
break;
--
Thanks for your comments.
I will post the next version with the suggested changes.
Regards,
Aboorva