Skip to contentSkip to conversation

ARM: dts: exynos: Add energy model for Exynos 5250

PATCH In review Danny Lin 2 participants 2 messages Jan 12, 2021 – Jan 13, 2021

Subsystems (2)
lore
Share & tools
lore

Danny Lin

author
#0 lore raw

PATCH 1/1 ARM: dts: exynos: Add energy model for Exynos 5250

This energy model enables the use of energy- and capacity-aware CPU frequency scaling.

Power and performance measurements were made using my freqbench [1] benchmark coordinator, which uses EEMBC CoreMark [2] as the workload and measures power usage using the integrated PMIC's fuel gauge (DS2784 in this case).

The energy model dynamic-power-coefficient values were calculated with
DPC = µW / MHz / V^2
for each OPP, and averaged across all OPPs within each cluster for the final coefficient.

A Google Nexus 10 device running a downstream 3.4 kernel was used for benchmarking to ensure proper frequency scaling and other low-level controls.

Raw benchmark results can be found in the freqbench repository [3]. Below is a human-readable summary:

===== CPU 1 =====
Frequencies: 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700

 200:   909     4.5 C/MHz    132 mW   36.2 J    6.9 I/mJ   275.0 s
 300:  1366     4.6 C/MHz    212 mW   38.7 J    6.5 I/mJ   183.0 s
 400:  1821     4.6 C/MHz    286 mW   39.3 J    6.4 I/mJ   137.3 s
 500:  2253     4.5 C/MHz    375 mW   41.7 J    6.0 I/mJ   111.0 s
 600:  2740     4.6 C/MHz    446 mW   40.7 J    6.1 I/mJ    91.2 s
 700:  3199     4.6 C/MHz    513 mW   40.1 J    6.2 I/mJ    78.2 s
 800:  3673     4.6 C/MHz    678 mW   46.1 J    5.4 I/mJ    68.1 s
 900:  4090     4.5 C/MHz    764 mW   46.7 J    5.4 I/mJ    61.1 s
1000:  4586     4.6 C/MHz    878 mW   47.9 J    5.2 I/mJ    54.5 s
1100:  5060     4.6 C/MHz   1084 mW   53.6 J    4.7 I/mJ    49.4 s
1200:  5515     4.6 C/MHz   1225 mW   55.5 J    4.5 I/mJ    45.3 s
1300:  5933     4.6 C/MHz   1396 mW   58.9 J    4.2 I/mJ    42.1 s
1400:  6395     4.6 C/MHz   1662 mW   65.0 J    3.8 I/mJ    39.1 s
1500:  6897     4.6 C/MHz   1895 mW   68.7 J    3.6 I/mJ    36.3 s
1600:  7332     4.6 C/MHz   2198 mW   75.0 J    3.3 I/mJ    34.1 s
1700:  7826     4.6 C/MHz   2497 mW   79.8 J    3.1 I/mJ    31.9 s

[1] https://github.com/kdrag0n/freqbench
[2] https://www.eembc.org/coremark/
[3] https://github.com/kdrag0n/freqbench/tree/master/results/exynos5250/main

---
 arch/arm/boot/dts/exynos5250.dtsi | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
arch/arm/boot/dts/exynos5250.dtsi +16 -0
index 2ea2caaca4e2..cc2fe0afcfc7 100644
@@ -58,6 +58,8 @@ cpu0: cpu@0 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;
+			dynamic-power-coefficient = <800>;
 		};
 		cpu1: cpu@1 {
 			device_type = "cpu";
@@ -67,6 +69,20 @@ cpu1: cpu@1 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;
+			dynamic-power-coefficient = <800>;
+		};
+
+		cpu-map {
+			cluster0 {
+				core0 {
+					cpu = <&cpu0>;
+				};
+
+				core1 {
+					cpu = <&cpu1>;
+				};
+			};
 		};
 	};
 
Signature · 7 lines
2.29.2


_______________________________________________
linux-arm-kernel mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
#1 lore raw
Quoted 7 lines · Danny Lin

This energy model enables the use of energy- and capacity-aware CPU frequency scaling.

Power and performance measurements were made using my freqbench [1] benchmark coordinator, which uses EEMBC CoreMark [2] as the workload and measures power usage using the integrated PMIC's fuel gauge (DS2784 in this case).

Thanks for the patch.

The tested setup is not representative here. The Nexus 10 is not supported by mainline and it might use specific revision of Exynos5250, different than present on other mainlined boards. One could assume that ratio of dynamic power coefficient of the cores should be similar... but there is no ratio here, as this is not a big.LITTLE platform.

Another problem is the v3.4 vendor kernel with a lot of out-of-tree code. This means it might use some different scheduler, different drivers, different voltages and a lot more of unknown differences. Vendor kernel should not matter that much in measurement of DPC but it makes the results not possible to reproduce.

You were also measuring the power delivered to entire system, no to CPUs, so you included static power in the data. Static power of CPUs and of the entire system.

Quoted 9 lines

The energy model dynamic-power-coefficient values were calculated with
DPC = µW / MHz / V^2
for each OPP, and averaged across all OPPs within each cluster for the final coefficient.

A Google Nexus 10 device running a downstream 3.4 kernel was used for benchmarking to ensure proper frequency scaling and other low-level controls.

Quoted 6 lines

Raw benchmark results can be found in the freqbench repository [3]. Below is a human-readable summary:

===== CPU 1 =====
Frequencies: 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500 1600 1700
200: 909 4.5 C/MHz 132 mW 36.2 J 6.9 I/mJ 275.0 s

What are the columns here? I would expect that fuel gauge gives you the current, but it's not here.

Quoted 30 lines
 300:  1366     4.6 C/MHz    212 mW   38.7 J    6.5 I/mJ   183.0 s
 400:  1821     4.6 C/MHz    286 mW   39.3 J    6.4 I/mJ   137.3 s
 500:  2253     4.5 C/MHz    375 mW   41.7 J    6.0 I/mJ   111.0 s
 600:  2740     4.6 C/MHz    446 mW   40.7 J    6.1 I/mJ    91.2 s
 700:  3199     4.6 C/MHz    513 mW   40.1 J    6.2 I/mJ    78.2 s
 800:  3673     4.6 C/MHz    678 mW   46.1 J    5.4 I/mJ    68.1 s
 900:  4090     4.5 C/MHz    764 mW   46.7 J    5.4 I/mJ    61.1 s
1000:  4586     4.6 C/MHz    878 mW   47.9 J    5.2 I/mJ    54.5 s
1100:  5060     4.6 C/MHz   1084 mW   53.6 J    4.7 I/mJ    49.4 s
1200:  5515     4.6 C/MHz   1225 mW   55.5 J    4.5 I/mJ    45.3 s
1300:  5933     4.6 C/MHz   1396 mW   58.9 J    4.2 I/mJ    42.1 s
1400:  6395     4.6 C/MHz   1662 mW   65.0 J    3.8 I/mJ    39.1 s
1500:  6897     4.6 C/MHz   1895 mW   68.7 J    3.6 I/mJ    36.3 s
1600:  7332     4.6 C/MHz   2198 mW   75.0 J    3.3 I/mJ    34.1 s
1700:  7826     4.6 C/MHz   2497 mW   79.8 J    3.1 I/mJ    31.9 s

[1] https://github.com/kdrag0n/freqbench
[2] https://www.eembc.org/coremark/
[3] https://github.com/kdrag0n/freqbench/tree/master/results/exynos5250/main

---
 arch/arm/boot/dts/exynos5250.dtsi | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
arch/arm/boot/dts/exynos5250.dtsi +0 -0
index 2ea2caaca4e2..cc2fe0afcfc7 100644

arch/arm/boot/dts/exynos5250.dtsi cpu0: cpu@0 { · line 61

Quoted 4 lines
arch/arm/boot/dts/exynos5250.dtsi +1 -0
@@ -58,6 +58,8 @@ cpu0: cpu@0 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;

The origin of this 1024 value should be explained.

arch/arm/boot/dts/exynos5250.dtsi cpu1: cpu@1 { · line 76

Quoted 13 lines
arch/arm/boot/dts/exynos5250.dtsi +6 -0
+			dynamic-power-coefficient = <800>;
 		};
 		cpu1: cpu@1 {
 			device_type = "cpu";
@@ -67,6 +69,20 @@ cpu1: cpu@1 {
 			clock-names = "cpu";
 			operating-points-v2 = <&cpu0_opp_table>;
 			#cooling-cells = <2>; /* min followed by max */
+			capacity-dmips-mhz = <1024>;
+			dynamic-power-coefficient = <800>;
+		};
+
+		cpu-map {

That's a second patch.

Best regards,
Krzysztof

Quoted 15 lines
arch/arm/boot/dts/exynos5250.dtsi +9 -0
+			cluster0 {
+				core0 {
+					cpu = <&cpu0>;
+				};
+
+				core1 {
+					cpu = <&cpu1>;
+				};
+			};
 		};
 	};
 
Signature · 1 line
2.29.2