[#QA-59] Unable to use the full CPU speed when CPUfreq Governor is ondemand.

Questions & Answers

Unable to use the full CPU speed when CPUfreq Governor is ondemand.

Created: 29/Sep/15 12:15 PM Updated: 02/Oct/15 01:45 PM

Fix Version/s:

None

File Attachments:

None

Image Attachments:

1. AWR.png
(49 kb)

2. EMTopActivity.png
(55 kb)

Product Version:	11.2.0.3
Operating System:	Linux
Operating System Version:	2.6.32-504.23.4.el6.x86_64
Host Name:	.
Database Name:	.

Description

« Hide

The customer is unable use the full CPU speed. The CPUfreq Governor is OnDemand.

All

Comments

Change History

Sort Order:

Ascending order - Click to sort in descending order

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 12:16 PM

See the following notes for the basic definitions of CPUfreq Governors:

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 12:21 PM

ENVIRONMENT:

Data:
for the CPU0(similar for the others):

[perftest1]/sys/devices/system/cpu/cpu0/cpufreq $ more *
::::::::::::::
affected_cpus
::::::::::::::
0
cpuinfo_cur_freq: Permission denied
::::::::::::::
cpuinfo_max_freq
::::::::::::::
2000000
::::::::::::::
cpuinfo_min_freq
::::::::::::::
1200000
::::::::::::::
cpuinfo_transition_latency
::::::::::::::
10000

*** ondemand: directory ***

::::::::::::::
related_cpus
::::::::::::::
0
::::::::::::::
scaling_available_frequencies
::::::::::::::
2000000 1900000 1800000 1700000 1600000 1500000 1400000 1300000 1200000
::::::::::::::
scaling_available_governors
::::::::::::::
ondemand userspace performance
::::::::::::::
scaling_cur_freq
::::::::::::::
2000000
::::::::::::::
scaling_driver
::::::::::::::
acpi-cpufreq
::::::::::::::
scaling_governor
::::::::::::::
ondemand
::::::::::::::
scaling_max_freq
::::::::::::::
2000000
::::::::::::::
scaling_min_freq
::::::::::::::
1200000
::::::::::::::
scaling_setspeed
::::::::::::::
<unsupported>

*** stats: directory ***

[perftest1]/sys/devices/system/cpu/cpu0/cpufreq $ cd ondemand
[perftest1]/sys/devices/system/cpu/cpu0/cpufreq/ondemand $ ls -ltr
total 0
-r--r--r-- 1 root root 4096 Sep 29 15:17 sampling_rate_min
-r--r--r-- 1 root root 4096 Sep 29 15:17 sampling_rate_max
-rw-r--r-- 1 root root 4096 Sep 29 15:17 up_threshold
-rw-r--r-- 1 root root 4096 Sep 29 15:17 sampling_rate
-rw-r--r-- 1 root root 4096 Sep 29 15:17 powersave_bias
-rw-r--r-- 1 root root 4096 Sep 29 15:17 ignore_nice_load
[perftest1]/sys/devices/system/cpu/cpu0/cpufreq/ondemand $ more *
::::::::::::::
ignore_nice_load
::::::::::::::
0
::::::::::::::
powersave_bias
::::::::::::::
0
::::::::::::::
sampling_rate
::::::::::::::
10000
::::::::::::::
sampling_rate_max
::::::::::::::
4294967295
::::::::::::::
sampling_rate_min
::::::::::::::
10000
::::::::::::::
up_threshold
::::::::::::::
95
[perftest1]/sys/devices/system/cpu/cpu0/cpufreq/ondemand $

View:

scaling_governor: CPU scaling governor is ondemand.
cpuinfo_min_freq: Minimum CPU frequency is 1200000Khz(1.2Ghz)
cpuinfo_max_freq: Maximum CPU frequency is 2001000Khz(2.0Ghz)
sampling_rate: The kernel looks at the CPU usage per 10000us(10ms) to make decisions about CPU frequency.
up_threshold: The kernel will increase the CPU frequency if average CPU usage between each sampling_rate(10ms) is higher than 95%.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 12:24 PM

atop(http://www.atoptool.nl/) tool wil be used to monitor CPU frequencies.

From the man page of atop:

In  case  that the kernel module 'cpufreq_stats' is active (after issueing 'modprobe cpufreq_stats'), the average frequency ('avgf')
            and the average scaling percentage ('avgscal') is shown. Otherwise the current frequency ('curf') and the current scaling percentage
            ('curscal') is shown at the moment that the sample is taken.

In order to compare the CPU usages to the frequencies, CPU "cpufreq_stats" should be enabled. Otherwise, atop will show the current frequencies, not the average during monitoring samples.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 12:29 PM

METHOD:

The tests will be done when CPU scaling governors are ondemand and then performance.
The same work load will be generated by HP's LOAD RUNNER tool.
The results will be compared.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 12:57 PM

TEST1:

CPU scaling governor is ondemand.

An atop snapshot:

ATOP - avsprddbflx05                2015/09/29  15:40:21                ---------                  10s elapsed
PRC | sys    7.86s | user  80.03s | #proc   1347 | #tslpi  1791 | #tslpu     0 | #zombie    0 | no  procacct |
CPU | sys      66% | user    800% | irq      13% | idle    626% | wait     95% | avgf 1.63GHz | avgscal  81% |
cpu | sys       4% | user     86% | irq       1% | idle      6% | cpu000 w  3% | avgf 1.94GHz | avgscal  96% |
cpu | sys       4% | user     77% | irq       5% | idle     11% | cpu004 w  3% | avgf 1.90GHz | avgscal  94% |
cpu | sys       4% | user     72% | irq       0% | idle     17% | cpu001 w  7% | avgf 1.83GHz | avgscal  91% |
cpu | sys       3% | user     67% | irq       0% | idle     20% | cpu002 w  9% | avgf 1.80GHz | avgscal  90% |
cpu | sys       3% | user     61% | irq       0% | idle     28% | cpu003 w  8% | avgf 1.73GHz | avgscal  86% |
cpu | sys       6% | user     54% | irq       1% | idle     34% | cpu009 w  6% | avgf 1.62GHz | avgscal  80% |
cpu | sys       3% | user     52% | irq       1% | idle     36% | cpu005 w  8% | avgf 1.68GHz | avgscal  83% |
cpu | sys       7% | user     47% | irq       1% | idle     29% | cpu008 w 16% | avgf 1.67GHz | avgscal  83% |
cpu | sys       6% | user     45% | irq       0% | idle     48% | cpu013 w  1% | avgf 1.52GHz | avgscal  76% |
cpu | sys       7% | user     39% | irq       1% | idle     53% | cpu015 w  1% | avgf 1.49GHz | avgscal  74% |
cpu | sys       3% | user     41% | irq       0% | idle     49% | cpu006 w  7% | avgf 1.57GHz | avgscal  78% |
cpu | sys       5% | user     34% | irq       0% | idle     52% | cpu010 w  9% | avgf 1.51GHz | avgscal  75% |
cpu | sys       2% | user     35% | irq       1% | idle     55% | cpu007 w  7% | avgf 1.55GHz | avgscal  77% |
cpu | sys       4% | user     32% | irq       0% | idle     63% | cpu014 w  1% | avgf 1.43GHz | avgscal  71% |
cpu | sys       4% | user     31% | irq       0% | idle     59% | cpu011 w  6% | avgf 1.46GHz | avgscal  72% |
cpu | sys       2% | user     26% | irq       0% | idle     68% | cpu012 w  3% | avgf 1.43GHz | avgscal  71% |
CPL | avg1    5.33 | avg5    5.46 | avg15   4.57 | csw   377039 | intr  323539 |              | numcpu    16 |
MEM | tot   126.1G | free   38.8G | cache   3.6G | dirty   4.0M | buff  146.3M | slab  577.8M |              |
SWP | tot    17.1G | free   17.1G |              |              |              | vmcom  12.9G | vmlim  42.6G |
NET | transport    | tcpi   20869 | tcpo   21016 | udpi   70875 | udpo   71067 | tcpao     33 | tcppo      1 |
NET | network      | ipi   128214 | ipo    92084 | ipfrw      0 | deliv  91742 | icmpi      0 | icmpo      0 |

  PID    TID   SYSCPU   USRCPU    VGROW   RGROW  RUID       EUID       THR  ST   EXC  S   CPU  CMD        1/64
13661      -    0.15s    4.50s   -24.0M  -14.3M  grid       oracle       1  --     -  R   47%  oracle
15747      -    0.16s    3.80s       0K    684K  grid       oracle       1  --     -  S   40%  oracle
13733      -    0.32s    3.57s   32768K  31360K  grid       oracle       1  --     -  R   39%  oracle
27274      -    0.61s    3.21s   24576K  11976K  grid       oracle       1  --     -  R   39%  oracle
14869      -    0.17s    3.29s       0K  -1880K  grid       oracle       1  --     -  S   35%  oracle

The "CPU" shows overall statistics for all CPUs.
The "cpu" shows statistics for single CPU.

Analysis:

Although maximum CPU frequency is 2.0Ghz, the server could not use its full speed. It used average 1.63Ghz, which is 81% of full CPU speed.
When CPU usage is 91%(sys:4+user:86+irq:1) at cpu000, it used average 1.94Ghz, which is 96% of full CPU speed.
When CPU usage is 51%(sys:6+user:45+irq:0) at cpu013, it used average 1.52Ghz, which is 76% of full CPU speed.
When CPU usage is 28%(sys:2+user:26+irq:0) at cpu012, it used average 1.43Ghz, which is 71% of full CPU speed.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 01:22 PM

TEST2:

CPU scaling governor is performance.

An atop snapshot:

ATOP - avsprddbflx05                2015/09/29  16:16:27                ---------                  10s elapsed
PRC | sys    7.06s | user  86.40s | #proc   1313 | #tslpi  1756 | #tslpu     0 | #zombie    0 | no  procacct |
CPU | sys      57% | user    864% | irq      14% | idle    623% | wait     43% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     80% | irq       7% | idle     10% | cpu004 w  1% | avgf 2.00GHz | avgscal 100% |
cpu | sys       5% | user     81% | irq       2% | idle     11% | cpu000 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     82% | irq       1% | idle     14% | cpu001 w  1% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     74% | irq       0% | idle     21% | cpu002 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     73% | irq       0% | idle     22% | cpu003 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     62% | irq       0% | idle     32% | cpu005 w  3% | avgf 2.00GHz | avgscal 100% |
cpu | sys       6% | user     58% | irq       1% | idle     26% | cpu008 w 10% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     51% | irq       0% | idle     43% | cpu006 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     47% | irq       0% | idle     47% | cpu010 w  3% | avgf 2.00GHz | avgscal 100% |
cpu | sys       3% | user     46% | irq       0% | idle     49% | cpu013 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       2% | user     44% | irq       1% | idle     50% | cpu007 w  3% | avgf 2.00GHz | avgscal 100% |
cpu | sys       6% | user     40% | irq       1% | idle     48% | cpu009 w  6% | avgf 2.00GHz | avgscal 100% |
cpu | sys       6% | user     33% | irq       1% | idle     57% | cpu011 w  3% | avgf 2.00GHz | avgscal 100% |
cpu | sys       2% | user     34% | irq       0% | idle     60% | cpu012 w  3% | avgf 2.00GHz | avgscal 100% |
cpu | sys       4% | user     28% | irq       0% | idle     66% | cpu014 w  2% | avgf 2.00GHz | avgscal 100% |
cpu | sys       2% | user     28% | irq       0% | idle     69% | cpu015 w  1% | avgf 2.00GHz | avgscal 100% |
CPL | avg1    5.98 | avg5    6.41 | avg15   4.75 | csw   382133 | intr  340254 |              | numcpu    16 |
MEM | tot   126.1G | free   36.6G | cache   5.3G | dirty  28.9M | buff  193.0M | slab  836.2M |              |
SWP | tot    17.1G | free   17.1G |              |              |              | vmcom  13.0G | vmlim  42.6G |
NET | transport    | tcpi   10272 | tcpo   10302 | udpi   75222 | udpo   75458 | tcpao     30 | tcppo      2 |
NET | network      | ipi   111030 | ipo    85760 | ipfrw      0 | deliv  85494 | icmpi      0 | icmpo      0 |

  PID    TID   SYSCPU   USRCPU    VGROW   RGROW  RUID       EUID       THR  ST   EXC  S   CPU  CMD        1/62
15847      -    0.17s    4.59s       0K    896K  grid       oracle       1  --     -  S   48%  oracle
14867      -    0.14s    3.79s       0K   2220K  grid       oracle       1  --     -  R   40%  oracle
15835      -    0.15s    3.76s    8192K    384K  grid       oracle       1  --     -  R   39%  oracle
14871      -    0.24s    3.59s       0K      0K  grid       oracle       1  --     -  R   39%  oracle
15849      -    0.14s    3.55s       0K  -1216K  grid       oracle       1  --     -  R   37%  oracle

Analysis:

The maximum CPU frequency is 2.0Ghz and all CPUs could use 100% of full CPU speed.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 03:19 PM

COMPARISION:

30 minutes load test results...

1st: When CPU scaling governor is ondemand.
2nd: When CPU scaling governor is performance.

Data:

Top Activity:

AWR:

Analysis:

"row cache lock" wait time decreased since the holders did their jobs faster, as a result held the resources shorter.
DB time decreased 31.4%, mostly from decrease in "row cache lock".
Logical reads increased 16.3% since more buffer gets could be done on the faster CPU frequency.

[ Permlink | « Hide ]

ubTools Support - 29/Sep/15 03:34 PM

SUMMARY:

Analysis:

Changing CPU scaling governor from "ondemand" to "performance" increased the performance.
Performance improvement is noticable when:
- The difference between the minumum and maximum CPU frequencies is high.
- CPU usage is not heavy(up_threshold:95%).
- There are sessions waiting for other sessions on CPU.

Recommendations:

If performance is important than heating, set CPU scaling governor to "performance".

[ Permlink | « Hide ]

ubTools Support - 30/Sep/15 09:24 AM

CPU TIME and LOGICAL READS:

Data:

	ondemand	performance	Difference(%)
CPU time per second	7.8s	8.3s	6.4
Logical reads per second	535,236.2	622,625.2	16.3
CPU time per Logical reads	14,6us	13,3us	8.9

Analysis:

8.9% improvements in CPU time caused 31.4% improvement DB time.

[ Permlink | « Hide ]

ubTools Support - 02/Oct/15 01:45 PM

The focus here is to show how CPU scaling governor affects Oracle service and wait times; not to show how to tune Oracle events such as "row cache lock" above.