sched/fair: add support to tune PELT ramp/decay timings

PATCH In review Patrick Bellasi 2 participants 2 messages Apr 9, 2018 – Apr 12, 2018

Subsystems (2)

Documentation Scheduler

lore

Share & tools

lore

Patrick Bellasi

author

2018-04-09 16:51 UTC #0 lore raw

PATCH 1/1 sched/fair: add support to tune PELT ramp/decay timings

The PELT half-life is the time [ms] required by the PELT signal to build up a 50% load/utilization, starting from zero. This time is currently hardcoded to be 32ms, a value which seems to make sense for most of the workloads.

However, 32ms has been verified to be too long for certain classes of workloads. For example, in the mobile space many tasks affecting the user-experience run with a 16ms or 8ms cadence, since they need to match the common 60Hz or 120Hz refresh rate of the graphics pipeline. This contributed so fare to the idea that "PELT is too slow" to properly track the utilization of interactive mobile workloads, especially compared to alternative load tracking solutions which provides a better representation of tasks demand in the range of 10-20ms.

A faster PELT ramp-up time could give some advantages to speed-up the time required for the signal to stabilize and thus to better represent task demands in the mobile space. As a downside, it also reduces the decay time, and thus we forget the load/utilization of sleeping tasks (or idle CPUs) faster.

Fortunately, since the integration of the utilization estimation support in mainline kernel:

commit 7f65ea42eb00 ("sched/fair: Add util_est on top of PELT")

a fast decay time is no longer an issue for tasks utilization estimation. Although estimated utilization does not slow down the decay of blocked utilization on idle CPUs, for mobile workloads this seems not to be a major concern compared to the benefits in interactivity responsiveness.

Let's add a compile time option to choose the PELT speed which better fits for a specific system. By default the current 32ms half-life is used, but we can also compile a kernel to use a faster ramp-up time of either 16ms or 8ms. These two configurations have been verified to give PELT a further improvement in performance, compared to other out-of-tree load tracking solutions, when it comes to track interactive workloads thus better supporting both tasks placements and frequencies selections.

Signed-off-by Patrick Bellasi <[email protected]>

Cc Ingo Molnar <[email protected]>

Cc Peter Zijlstra <[email protected]>

Cc Jonathan Corbet <[email protected]>

Cc Paul Turner <[email protected]>

Cc Vincent Guittot <[email protected]>

Cc Joel Fernandes <[email protected]>

Cc Morten Rasmussen <[email protected]>

Cc [email protected]

---
 Documentation/scheduler/sched-pelt.c | 45 ++++++++++++++++++++++--------------
 init/Kconfig                         | 44 +++++++++++++++++++++++++++++++++++
 kernel/sched/sched-pelt.h            | 39 ++++++++++++++++++++++++++-----
 3 files changed, 105 insertions(+), 23 deletions(-)

Documentation/scheduler/sched-pelt.c +28 ~~-17~~

index e4219139386a..e0ae21616188 100644
@@ -10,34 +10,35 @@
 #include <math.h>
 #include <stdio.h>
 
-#define HALFLIFE 32
+#define HALFLIFE { 32, 16, 8 }
 #define SHIFT 32
 
 double y;
 
-void calc_runnable_avg_yN_inv(void)
+void calc_runnable_avg_yN_inv(const int halflife)
 {
 	int i;
 	unsigned int x;
 
 	printf("static const u32 runnable_avg_yN_inv[] = {");
-	for (i = 0; i < HALFLIFE; i++) {
+	for (i = 0; i < halflife; i++) {
 		x = ((1UL<<32)-1)*pow(y, i);
 
-		if (i % 6 == 0) printf("\n\t");
-		printf("0x%8x, ", x);
+		if (i % 4 == 0)
+			printf("\n\t");
+		printf("0x%8x,", x);
 	}
 	printf("\n};\n\n");
 }
 
 int sum = 1024;
 
-void calc_runnable_avg_yN_sum(void)
+void calc_runnable_avg_yN_sum(const int halflife)
 {
 	int i;
 
 	printf("static const u32 runnable_avg_yN_sum[] = {\n\t    0,");
-	for (i = 1; i <= HALFLIFE; i++) {
+	for (i = 1; i <= halflife; i++) {
 		if (i == 1)
 			sum *= y;
 		else

@@ -55,7 +56,7 @@ int n = -1;
 /* first period */
 long max = 1024;
 
-void calc_converged_max(void)
+void calc_converged_max(const int halflife)
 {
 	long last = 0, y_inv = ((1UL<<32)-1)*y;

@@ -73,17 +74,17 @@ void calc_converged_max(void)
 		last = max;
 	}
 	n--;
-	printf("#define LOAD_AVG_PERIOD %d\n", HALFLIFE);
+	printf("#define LOAD_AVG_PERIOD %d\n", halflife);
 	printf("#define LOAD_AVG_MAX %ld\n", max);
-//	printf("#define LOAD_AVG_MAX_N %d\n\n", n);
+	/* printf("#define LOAD_AVG_MAX_N %d\n\n", n); */
 }
 
-void calc_accumulated_sum_32(void)
+void calc_accumulated_sum_32(const int halflife)
 {
 	int i, x = sum;
 
 	printf("static const u32 __accumulated_sum_N32[] = {\n\t     0,");
-	for (i = 1; i <= n/HALFLIFE+1; i++) {
+	for (i = 1; i <= n / halflife + 1; i++) {
 		if (i > 1)
 			x = x/2 + sum;

@@ -97,12 +98,22 @@ void calc_accumulated_sum_32(void)
 
 void main(void)
 {
+	int hl_value[] = HALFLIFE;
+	int hl_count = sizeof(hl_value) / sizeof(int);
+	int hl_idx, halflife;
+
 	printf("/* Generated by Documentation/scheduler/sched-pelt; do not modify. */\n\n");
 
-	y = pow(0.5, 1/(double)HALFLIFE);
+	for (hl_idx = 0; hl_idx < hl_count; ++hl_idx) {
+		halflife = hl_value[hl_idx];
 
-	calc_runnable_avg_yN_inv();
-//	calc_runnable_avg_yN_sum();
-	calc_converged_max();
-//	calc_accumulated_sum_32();
+		y = pow(0.5, 1 / (double)halflife);
+
+		printf("\n#ifdef CONFIG_PELT_HALFLIFE_%d\n", halflife);
+		calc_runnable_avg_yN_inv(halflife);
+		/* calc_runnable_avg_yN_sum(halflife); */
+		calc_converged_max(halflife);
+		/* calc_accumulated_sum_32(halflife); */
+		printf("#endif\n");
+	}
 }

init/Kconfig +44 -0

50 lines not on this page · raw message

kernel/sched/sched-pelt.h +33 -6

index a26473674fb7..c978fe03f788 100644
@@ -1,14 +1,41 @@
 /* SPDX-License-Identifier: GPL-2.0 */
 /* Generated by Documentation/scheduler/sched-pelt; do not modify. */
 
+#ifdef CONFIG_PELT_HALFLIFE_32
 static const u32 runnable_avg_yN_inv[] = {
-	0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 0xeac0c6e6, 0xe5b906e6,
-	0xe0ccdeeb, 0xdbfbb796, 0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85,
-	0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 0xb504f333, 0xb123f581,
-	0xad583ee9, 0xa9a15ab4, 0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9,
-	0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 0x8b95c1e3, 0x88980e80,
-	0x85aac367, 0x82cd8698,
+	0xffffffff, 0xfa83b2da, 0xf5257d14, 0xefe4b99a, 
+	0xeac0c6e6, 0xe5b906e6, 0xe0ccdeeb, 0xdbfbb796, 
+	0xd744fcc9, 0xd2a81d91, 0xce248c14, 0xc9b9bd85, 
+	0xc5672a10, 0xc12c4cc9, 0xbd08a39e, 0xb8fbaf46, 
+	0xb504f333, 0xb123f581, 0xad583ee9, 0xa9a15ab4, 
+	0xa5fed6a9, 0xa2704302, 0x9ef5325f, 0x9b8d39b9, 
+	0x9837f050, 0x94f4efa8, 0x91c3d373, 0x8ea4398a, 
+	0x8b95c1e3, 0x88980e80, 0x85aac367, 0x82cd8698, 
 };
 
 #define LOAD_AVG_PERIOD 32
 #define LOAD_AVG_MAX 47742
+#endif
+
+#ifdef CONFIG_PELT_HALFLIFE_16
+static const u32 runnable_avg_yN_inv[] = {
+	0xffffffff, 0xf5257d14, 0xeac0c6e6, 0xe0ccdeeb, 
+	0xd744fcc9, 0xce248c14, 0xc5672a10, 0xbd08a39e, 
+	0xb504f333, 0xad583ee9, 0xa5fed6a9, 0x9ef5325f, 
+	0x9837f050, 0x91c3d373, 0x8b95c1e3, 0x85aac367, 
+};
+
+#define LOAD_AVG_PERIOD 16
+#define LOAD_AVG_MAX 24152
+#endif
+
+#ifdef CONFIG_PELT_HALFLIFE_8
+static const u32 runnable_avg_yN_inv[] = {
+	0xffffffff, 0xeac0c6e6, 0xd744fcc9, 0xc5672a10, 
+	0xb504f333, 0xa5fed6a9, 0x9837f050, 0x8b95c1e3, 
+};
+
+#define LOAD_AVG_PERIOD 8
+#define LOAD_AVG_MAX 12337
+#endif
+

Signature · 1 line

2.15.1

Peter Zijlstra

maintainer

2018-04-12 19:13 UTC #1 lore raw

Quoted 5 lines · Patrick Bellasi

The PELT half-life is the time [ms] required by the PELT signal to build up a 50% load/utilization, starting from zero. This time is currently hardcoded to be 32ms, a value which seems to make sense for most of the workloads.

Quoted 8 lines

However, 32ms has been verified to be too long for certain classes of workloads. For example, in the mobile space many tasks affecting the user-experience run with a 16ms or 8ms cadence, since they need to match the common 60Hz or 120Hz refresh rate of the graphics pipeline. This contributed so fare to the idea that "PELT is too slow" to properly track the utilization of interactive mobile workloads, especially compared to alternative load tracking solutions which provides a better representation of tasks demand in the range of 10-20ms.

Initially the 32 was chosen to more or less correspond to the effective scheduling period (sysctl_sched_latency based). The thinking was that if you pick a PELT window shorter than the period, the result becomes unstable due to not all tasks getting an equal go at things.

(of course, stuffing enough tasks on a rq will break this, but at that point you have worse problems to deal with)

Should we retain this? Esp. with the lower end (8ms) I worry we'll see more of those effects.

Quoted 9 lines

Fortunately, since the integration of the utilization estimation support in mainline kernel:
commit 7f65ea42eb00 ("sched/fair: Add util_est on top of PELT")
a fast decay time is no longer an issue for tasks utilization estimation. Although estimated utilization does not slow down the decay of blocked utilization on idle CPUs, for mobile workloads this seems not to be a major concern compared to the benefits in interactivity responsiveness.

By picking a smaller PELT window, the util_est window shrinks correspondingly; is that intentional or do we want to modify UTIL_EST_WEIGHT_SHIFT to negate the PELT window changes?
--
To unsubscribe from this list: send the line "unsubscribe linux-doc" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html