LinuxLists.cc - [intel_pstate] fcb6a15c: -34.4% aim7 jobs-per-min and -17.0% will-it-scale/page_fault1 Pkg

2014-02-07 12:33:26

Subject: [intel_pstate] fcb6a15c: -34.4% aim7 jobs-per-min and -17.0% will-it-scale/page_fault1 Pkg_W

Hi Dirk,

FYI, we notice dropped power and performance numbers on commit
fcb6a15c2e7e76d493e6f91ea889ab40e1c643a4 ("intel_pstate: Take core C0
time into account for core busy calculation").

v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
--------------- -------------------------
6.53 ~ 4% -12.0% 5.75 ~ 2% lkp-sb03/micro/netperf/120s-200%-SCTP_STREAM
147.93 ~ 0% -23.3% 113.40 ~ 0% lkp-sbx04/micro/will-it-scale/page_fault1
154.46 ~ 0% -22.9% 119.15 ~ 0% TOTAL turbostat.Cor_W

v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
--------------- -------------------------
203 ~ 0% -17.0% 169 ~ 0% lkp-sbx04/micro/will-it-scale/page_fault1
203 ~ 0% -17.0% 169 ~ 0% TOTAL turbostat.Pkg_W

v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
--------------- -------------------------
572540 ~ 1% -30.2% 399351 ~ 1% grantley/micro/aim7/disk_src
317151 ~ 1% -42.0% 183863 ~ 0% grantley/micro/aim7/shell_rtns_1
889692 ~ 1% -34.4% 583214 ~ 1% TOTAL aim7.2000.jobs-per-min

More details are attached.

time.user_time

600 ++-------------------------------------------------------------------+
O O O O O |
500 ++O O O O O O O O O O O O O O O O O O O O O O O O
| |
| |
400 ++ |
| |
300 ++ |
| |
200 ++ |
| |
| |
100 ++ |
| |
0 *+*--*-*--*-*--*-*--*-*--*-*--*-*--*-*-*--*--------------------------+

time.system_time

1200 ++------------------------------------------------------------------+
| |
1000 O+O O O O O O O O O O O O O O O O O O O O O O O O O O O O
| |
| |
800 ++ |
| |
600 ++ |
| |
400 ++ |
| |
| |
200 ++ |
*.*..*.*..*.*..*.*.*..*.*..*.*..*.*.*..*.* |
0 ++------------------------------------------------------------------+

time.percent_of_cpu_this_job_got

2500 O+O--O-O--O------O------O--O----O-O-O--O-O--O-O-----------O---------+
| O O O O O O O O O O O O O
| |
2000 ++ |
| |
| |
1500 ++ |
| |
1000 ++ |
| |
| |
500 ++ |
| |
*.*..*.*..*.*..*.*.*..*.*..*.*..*.*.*..*.* |
0 ++------------------------------------------------------------------+

time.elapsed_time

70 ++----------------------------O---------------------------------------+
| O O |
65 O+ O O O O O O O O O O O O O O O O O O O O O O
| O O O |
60 ++ |
| |
55 ++ |
| |
50 ++ |
| |
45 ++ |
| |
40 ++ .*. .*. .*. |
*..*.*. *..*.*..*.*. *..*.*..*.*. *..* |
35 ++--------------------------------------------------------------------+

aim7.2000.jobs-per-min

340000 ++----------------------------------------------------------------+
|.*.. .*.. .*.. .* *.* |
320000 *+ *. .*..* *.*. .*.* * + .. |
300000 ++ * *. * |
| |
280000 ++ |
| |
260000 ++ |
| |
240000 ++ |
220000 ++ |
| |
200000 ++ |
O O O O O O O O O |
180000 ++-------O--O-O----O-O----O----O-O-O--O------O-O--O-O-O--O-O-O--O-O

Thanks,
Fengguang

Attachments:

(No filename) (7.09 kB)
fcb6a15c2e7e76d493e6f91ea889ab40e1c643a4 (276.05 kB)
Download all attachments

2014-02-10 15:54:25

by Dirk Brandewie

[permalink] [raw]

Subject: Re: [intel_pstate] fcb6a15c: -34.4% aim7 jobs-per-min and -17.0% will-it-scale/page_fault1 Pkg_W

On 02/07/2014 04:33 AM, Fengguang Wu wrote:
> Hi Dirk,
>
> FYI, we notice dropped power and performance numbers on commit
> fcb6a15c2e7e76d493e6f91ea889ab40e1c643a4 ("intel_pstate: Take core C0
> time into account for core busy calculation").
>
>
> v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
> --------------- -------------------------
> 6.53 ~ 4% -12.0% 5.75 ~ 2% lkp-sb03/micro/netperf/120s-200%-SCTP_STREAM
> 147.93 ~ 0% -23.3% 113.40 ~ 0% lkp-sbx04/micro/will-it-scale/page_fault1
> 154.46 ~ 0% -22.9% 119.15 ~ 0% TOTAL turbostat.Cor_W
>
> v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
> --------------- -------------------------
> 203 ~ 0% -17.0% 169 ~ 0% lkp-sbx04/micro/will-it-scale/page_fault1
> 203 ~ 0% -17.0% 169 ~ 0% TOTAL turbostat.Pkg_W
>

These two tests show an improvement in power :-)

> v3.14-rc1 fcb6a15c2e7e76d493e6f91ea
> --------------- -------------------------
> 572540 ~ 1% -30.2% 399351 ~ 1% grantley/micro/aim7/disk_src
> 317151 ~ 1% -42.0% 183863 ~ 0% grantley/micro/aim7/shell_rtns_1
> 889692 ~ 1% -34.4% 583214 ~ 1% TOTAL aim7.2000.jobs-per-min
>

I am investigating where the aim benchmark went south.