[PlanetCCRMA] Performance degradation with kernel-rt-3.14.17-200.rt9.1.fc20.ccrma.x86_64

Oleg Samarin osamarin68 at gmail.com
Thu Sep 11 10:49:14 PDT 2014


В Ср, 10/09/2014 в 23:01 -0700, Fernando Lopez-Lezcano пишет:
> On 09/09/2014 09:57 PM, Oleg Samarin wrote:
> > I've made more testing
> >
> >>> echo "100" > /sys/devices/system/cpu/intel_pstate/min_perf_pct increases
> >>> CPU does not speed up so much so it cann't be a workaround.
> >>
> >> How much faster does it go with min_perf_pct set to 100?
> >
> > Seems, it solves the issue:
> >
> > 3.4.17, min_perf_pct=100:
> >          Core [core-id]  :Actual Freq (Mult.)	  C0%   Halt(C1)%  C3 %   C6 %
> >          Core 1 [0]:	  4420.54 (44.21x)      26.8      56    8.22       1
> >          Core 2 [1]:	  4483.19 (44.83x)      49.9     8.1    25.2       1
> >          Core 3 [2]:	  4509.71 (45.10x)      9.28    21.2    65.5       1
> >          Core 4 [3]:	  4559.41 (45.59x)      57.7    1.32    19.1    2.19
> 
> Glad to hear that!
> That is my experience as well...
> 

Some more test results:

kernel-rt.3.12.12 with intel_pstate from trunk:

Core [core-id]  :Actual Freq (Mult.)	  C0%   Halt(C1)%  C3 %   C6 %
Core 1 [0]:	  1634.24 (16.34x)      15.7    84.8        6.7       1
Core 2 [1]:	  1648.80 (16.49x)      9.76    31.3    62.2    1.73
Core 3 [2]:	  1630.97 (16.31x)      88.7    50.1    2.66    4.72
Core 4 [3]:	  1634.79 (16.35x)      40.9    66.7    9.51     4.1

kernel-rt.3.14.17 with intel_pstate from 3.12.12        
Core [core-id]  :Actual Freq (Mult.)	  C0%   Halt(C1)%  C3 %   C6 %
Core 1 [0]:	  4531.24 (45.31x)      27.8    20.9      41       1
Core 2 [1]:	  4533.55 (45.34x)      17.5    35.8    40.9       0
Core 3 [2]:	  4550.54 (45.51x)      42.4    20.7    21.3    1.29
Core 4 [3]:	  4534.97 (45.35x)      27.3    60.6       1    2.06

So the new behaviour of intel_pstate driver is the only cause of the
performance problems in 3.14.17

> Another one to try is no_turbo - that would get you a more "even" speed 
> but you loose speed when running processes on few cpus. It performed 
> worse than no_turbo=0 in my limited tests.
> 

It prevents growing the CPU speed above 3.4 Ghz. It is not good, because
the maximum turbo speed of my CPU is 4.8 Ghz:

kernel-rt.3.12.12 with intel_pstate from trunk and no_turbo=1

Core [core-id]  :Actual Freq (Mult.)	  C0%   Halt(C1)%  C3 %   C6 %
Core 1 [0]:	  3373.71 (33.75x)        16    53.9    29.2       1
Core 2 [1]:	  3386.71 (33.88x)        19      46      34       1
Core 3 [2]:	  3389.96 (33.91x)      13.6    61.3    23.1    2.06
Core 4 [3]:	  3380.79 (33.82x)      10.1    77.3    8.88    3.71




More information about the PlanetCCRMA mailing list