The similar test run again against pad-gate head 8101. I found in idle p-state
transition times was reduced, but the number isn't improved when the system
under a workload. The data as folllows:
avg workload pstate change times/minutes
(1) (2) Avg
=========================================
0% 72 24 48
13% 3696 3784 3740
27% 6544 6712 6628
40% 9512 9536 9524
53% 11152 11144 11148
65% 12648 12296 12472
77% 11008 11800 11404
88% 8888 8776 8832
100% 8656 8740 8698
==========================================
>From the data, We see the pstate transition still occuring frequently when the
>system is
under a max load. That can explain why we found the peak throughput has a
regression
in SPECpower report under event mode.
Thanks,
-Aubrey
Eric.Saxe wrote:
> Hi Aubrey,
>
> Great data. Can you send me the workload? I'd like to reproduce to see
> where/how things are going wrong.
>
> Thanks!
> -Eric
>
> Li, Aubrey wrote:
>> Testing with a graduated workload shows how well event mode works.
>>
>> cpupm is set to event-mode,
>> c-state is disabled,
>> the hardware platfrom has 8 cores.
>> monitor dtrace probe "ess_transition"
>>
>> The result is as follows:
>> =========================================
>> average | the numbers of the probe
>> system workload | is hitted per minutes(60s)
>>- - - - - - - - - - - - - - - - - - - - -
>> 0% | 243
>> 13% | 28195
>> 26% | 5379
>> 38% | 8723
>> 51% | 9939
>> 63% | 12259
>> 76% | 10611
>> 88% | 11907
>> 100% | 13443
>> ========================================
>> The average system workload was obtained from (100 - vmstat.idle).
>>> From the result, we can see, event-mode is too sensitive to
> properly drive
>> the p-state transition. this could be a reason to impact performance.
>>
>> Any thoughts?
>>
>> Thanks,
>> -Aubrey
>>
>>
>>
>> _______________________________________________
>> tesla-dev mailing list
>> tesla-dev at opensolaris.org
>> http://mail.opensolaris.org/mailman/listinfo/tesla-dev
