使用cpufreq-bench评估cpufreq策略对系统性能的影响

Linux 3.1已经将cpupower-utils中放入内核的tools目录，见：

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=tree;f=tools/power;h=f6a98ba17528ed7645f53bb8cb67f59112adf346;hb=c3b92c8787367a8bb53d57d9789b558f1295cc96

在其中有一个工具叫cpufreq-bench，可以用于测试ondemand等cpufreq策略的使用对系统性能的影响，譬如：

# /usr/sbin/cpufreq-bench -l 50000 -s 100000 -x 50000 -y 100000 -g ondemand -r 5 -n 5 -v

loading defaults

parsing: sleep -> 50000

parsing: load -> 50000

parsing: cpu -> 0

parsing: priority -> LOW

parsing: output -> /var/log/cpufreq-bench

logilename: /var/log/cpufreq-bench/benchmark_localhost_2.6.38.8-xxx_1686167.log

Logfile: /var/log/cpufreq-bench/benchmark_localhost_2.6.38.8-xxx_1686167.log

parsing: sleep_step -> 50000

parsing: load_step -> 50000

parsing: cycles -> 20

parsing: rounds -> 40

parsing: verbose -> 0

parsing: governor -> ondemand

user load time -> 50000

user sleep time -> 100000

user load_step -> 50000

user sleep_step -> 100000

user governor -> ondemand

user rounds -> 5

user cycles -> 5

verbose output enabled

starting benchmark with parameters:

config:

        sleep=100000

        load=50000

        sleep_step=100000

        load_step=50000

        cpu=0

        cycles=5

        rounds=5

        governor=ondemand

 

approx. test duration: 0m

set cpu affinity to cpu #0

set affinity to cpu #0

low priority condition requested

set scheduler priority to 0

set performance as cpufreq governor

calibrating load of 50000us, please wait...

calibrating with 6 rounds

calibrating with 6 rounds

calibrating with 6 rounds

calibrating with 6 rounds

calibration done

_round 1: doing 5 cycles with 6 calculations for 50000us

avarage: 8333us, rps:120

performance cycle took 147613us, sleep: 100000us, load: 50000us, rounds: 6

performance cycle took 211700us, sleep: 100000us, load: 50000us, rounds: 6

performance cycle took 147747us, sleep: 100000us, load: 50000us, rounds: 6

performance cycle took 147537us, sleep: 100000us, load: 50000us, rounds: 6

performance cycle took 147482us, sleep: 100000us, load: 50000us, rounds: 6

set ondemand as cpufreq governor

powersave cycle took 149627us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took 243281us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took 288620us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took 288554us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took 290043us, sleep: 100000us, load: 50000us, rounds: 6

performance is at 39.74%

set performance as cpufreq governor

calibrating load of 100000us, please wait...

calibrating with 12 rounds

calibrating with 12 rounds

calibrating with 12 rounds

calibrating with 12 rounds

calibration done

_round 2: doing 5 cycles with 12 calculations for 100000us

avarage: 8333us, rps:120

performance cycle took 296126us, sleep: 200000us, load: 100000us, rounds: 12

performance cycle took 296454us, sleep: 200000us, load: 100000us, rounds: 12

performance cycle took 296068us, sleep: 200000us, load: 100000us, rounds: 12

performance cycle took 296026us, sleep: 200000us, load: 100000us, rounds: 12

performance cycle took 295887us, sleep: 200000us, load: 100000us, rounds: 12

set ondemand as cpufreq governor

powersave cycle took 479989us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took 464829us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took 389619us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took 514207us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took 473326us, sleep: 200000us, load: 100000us, rounds: 12

performance is at 36.35%

set performance as cpufreq governor

calibrating load of 150000us, please wait...

calibrating with 18 rounds

calibrating with 18 rounds

calibrating with 18 rounds

calibrating with 18 rounds

calibration done

_round 3: doing 5 cycles with 18 calculations for 150000us

avarage: 8333us, rps:120

performance cycle took 445176us, sleep: 300000us, load: 150000us, rounds: 18

performance cycle took 445219us, sleep: 300000us, load: 150000us, rounds: 18

performance cycle took 444955us, sleep: 300000us, load: 150000us, rounds: 18

performance cycle took 445073us, sleep: 300000us, load: 150000us, rounds: 18

performance cycle took 445084us, sleep: 300000us, load: 150000us, rounds: 18

set ondemand as cpufreq governor

powersave cycle took 563155us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took 598616us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took 614820us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took 618511us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took 619225us, sleep: 300000us, load: 150000us, rounds: 18

performance is at 47.91%

set performance as cpufreq governor

calibrating load of 200000us, please wait...

calibrating with 24 rounds

calibrating with 25 rounds

calibrating with 25 rounds

calibrating with 25 rounds

calibration done

_round 4: doing 5 cycles with 25 calculations for 200000us

avarage: 8000us, rps:125

performance cycle took 602496us, sleep: 400000us, load: 200000us, rounds: 25

performance cycle took 602759us, sleep: 400000us, load: 200000us, rounds: 25

performance cycle took 602782us, sleep: 400000us, load: 200000us, rounds: 25

performance cycle took 602688us, sleep: 400000us, load: 200000us, rounds: 25

performance cycle took 602433us, sleep: 400000us, load: 200000us, rounds: 25

set ondemand as cpufreq governor

powersave cycle took 756766us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took 774079us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took 778619us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took 779616us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took 779655us, sleep: 400000us, load: 200000us, rounds: 25

performance is at 54.22%

set performance as cpufreq governor

calibrating load of 250000us, please wait...

calibrating with 30 rounds

calibrating with 31 rounds

calibrating with 31 rounds

calibrating with 31 rounds

calibration done

_round 5: doing 5 cycles with 31 calculations for 250000us

avarage: 8064us, rps:124

performance cycle took 752761us, sleep: 500000us, load: 250000us, rounds: 31

performance cycle took 751929us, sleep: 500000us, load: 250000us, rounds: 31

performance cycle took 752054us, sleep: 500000us, load: 250000us, rounds: 31

performance cycle took 751597us, sleep: 500000us, load: 250000us, rounds: 31

performance cycle took 751791us, sleep: 500000us, load: 250000us, rounds: 31

set ondemand as cpufreq governor

powersave cycle took 958276us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took 928743us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took 921301us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took 920670us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took 920122us, sleep: 500000us, load: 250000us, rounds: 31

performance is at 58.64%

 

上述测试结果可以看出，被测试平台ondemand采用后，性能下降很大：

Round 1 - 39.74%

Round 2 - 36.35%

Round 3 - 47.91%

Round 4 - 54.22%

Round 5 - 58.64%

比较理想的结果是，ondemand采用后性能维持在performance策略的90％左右：

cpufreq-bench results with optimization on OMAP:

Round 1 - 90.24%

Round 2 - 94.48%

Round 3 - 96.06%

Round 4 - 96.6%

Round 5 - 86.89%

 

cpufreq-bench results on X86 platform:

Round 1 - 88.67%

Round 2 - 94.71%

Round 3 - 95.53%

Round 4 - 96.34%

Round 5 - 98.03%

由此可见，被测平台的cpufreq驱动相关参数还需要深度调整。


我们在同样的平台上测试Android添加的interactive这个cpufreq策略的性能：

#/usr/sbin/cpufreq-bench -l 50000 -s 100000 -x 50000 -y 100000 -g interactive -

r 5 -n 5 -v

loading defaults

parsing: sleep ->50000

parsing: load ->50000

parsing: cpu -> 0

parsing: priority-> LOW

parsing: output ->/var/log/cpufreq-bench

logilename:/var/log/cpufreq-bench/benchmark_localhost_2.6.38.8-xxx_6470.log

Logfile:/var/log/cpufreq-bench/benchmark_localhost_2.6.38.8-xxx_6470.log

parsing: sleep_step-> 50000

parsing: load_step-> 50000

parsing: cycles ->20

parsing: rounds ->40

parsing: verbose-> 0

parsing: governor-> ondemand

user load time ->50000

user sleep time ->100000

user load_step ->50000

user sleep_step ->100000

user governor ->interactive

user rounds -> 5

user cycles -> 5

verbose outputenabled

starting benchmarkwith parameters:

config:

        sleep=100000

        load=50000

        sleep_step=100000

        load_step=50000

        cpu=0

        cycles=5

        rounds=5

        governor=interactive

 

approx. testduration: 0m

set cpu affinity tocpu #0

set affinity to cpu#0

low prioritycondition requested

set schedulerpriority to 0

set performance ascpufreq governor

calibrating load of50000us, please wait...

calibrating with 6rounds

calibrating with 6rounds

calibrating with 6rounds

calibrating with 6rounds

calibration done

_round 1: doing 5cycles with 6 calculations for 50000us

avarage: 8333us,rps:120

performance cycletook 147635us, sleep: 100000us, load: 50000us, rounds: 6

performance cycletook 147506us, sleep: 100000us, load: 50000us, rounds: 6

performance cycletook 147590us, sleep: 100000us, load: 50000us, rounds: 6

performance cycletook 147755us, sleep: 100000us, load: 50000us, rounds: 6

performance cycletook 147563us, sleep: 100000us, load: 50000us, rounds: 6

set interactive ascpufreq governor

powersave cycle took166866us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took164004us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took165926us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took163750us, sleep: 100000us, load: 50000us, rounds: 6

powersave cycle took165789us, sleep: 100000us, load: 50000us, rounds: 6

performance is at 72.95%

set performance ascpufreq governor

calibrating load of100000us, please wait...

calibrating with 12rounds

calibrating with 12rounds

calibrating with 12rounds

calibrating with 12rounds

calibration done

_round 2: doing 5cycles with 12 calculations for 100000us

avarage: 8333us,rps:120

performance cycletook 296382us, sleep: 200000us, load: 100000us, rounds: 12

performance cycletook 296349us, sleep: 200000us, load: 100000us, rounds: 12

performance cycletook 296266us, sleep: 200000us, load: 100000us, rounds: 12

performance cycletook 296152us, sleep: 200000us, load: 100000us, rounds: 12

performance cycletook 295848us, sleep: 200000us, load: 100000us, rounds: 12

set interactive ascpufreq governor

powersave cycle took314330us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took308119us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took309546us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took309942us, sleep: 200000us, load: 100000us, rounds: 12

powersave cycle took309673us, sleep: 200000us, load: 100000us, rounds: 12

performance is at 87.20%

set performance ascpufreq governor

calibrating load of150000us, please wait...

calibrating with 18rounds

calibrating with 18rounds

calibrating with 18rounds

calibrating with 18rounds

calibration done

_round 3: doing 5cycles with 18 calculations for 150000us

avarage: 8333us,rps:120

performance cycletook 445460us, sleep: 300000us, load: 150000us, rounds: 18

performance cycletook 445368us, sleep: 300000us, load: 150000us, rounds: 18

performance cycletook 445724us, sleep: 300000us, load: 150000us, rounds: 18

performance cycletook 445194us, sleep: 300000us, load: 150000us, rounds: 18

performance cycletook 445266us, sleep: 300000us, load: 150000us, rounds: 18

set interactive ascpufreq governor

powersave cycle took457701us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took462000us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took457623us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took461664us, sleep: 300000us, load: 150000us, rounds: 18

powersave cycle took458122us, sleep: 300000us, load: 150000us, rounds: 18

performance is at 91.21%

set performance ascpufreq governor

calibrating load of200000us, please wait...

calibrating with 24rounds

calibrating with 25rounds

calibrating with 25rounds

calibrating with 25rounds

calibration done

_round 4: doing 5cycles with 25 calculations for 200000us

avarage: 8000us,rps:125

performance cycletook 602273us, sleep: 400000us, load: 200000us, rounds: 25

performance cycletook 602022us, sleep: 400000us, load: 200000us, rounds: 25

performance cycletook 602541us, sleep: 400000us, load: 200000us, rounds: 25

performance cycletook 602642us, sleep: 400000us, load: 200000us, rounds: 25

performance cycletook 603079us, sleep: 400000us, load: 200000us, rounds: 25

set interactive ascpufreq governor

powersave cycle took616433us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took613765us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took616331us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took614218us, sleep: 400000us, load: 200000us, rounds: 25

powersave cycle took615348us, sleep: 400000us, load: 200000us, rounds: 25

performance is at 94.10%

set performance ascpufreq governor

calibrating load of250000us, please wait...

calibrating with 30rounds

calibrating with 31rounds

calibrating with 31rounds

calibrating with 31rounds

calibration done

_round 5: doing 5cycles with 31 calculations for 250000us

avarage: 8064us,rps:124

performance cycletook 752050us, sleep: 500000us, load: 250000us, rounds: 31

performance cycletook 751701us, sleep: 500000us, load: 250000us, rounds: 31

performance cycletook 751966us, sleep: 500000us, load: 250000us, rounds: 31

performance cycletook 752029us, sleep: 500000us, load: 250000us, rounds: 31

performance cycletook 752338us, sleep: 500000us, load: 250000us, rounds: 31

set interactive ascpufreq governor

powersave cycle took768047us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took763706us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took766064us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took763898us, sleep: 500000us, load: 250000us, rounds: 31

powersave cycle took765694us, sleep: 500000us, load: 250000us, rounds: 31

performance is at 94.93%

可见它的结果为：

Round 1 - 72.95%

Round 2 - 87.20%

Round 3 - 91.21%

Round 4 - 94.10%

Round 5 - 94.93%

interactive为UI或者说IO更好的响应而设计，测试结果也显示它在性能方面的考量更加aggressive一些。