Rafael Vanoni wrote:
> Mark Haywood wrote:
>> Li, Aubrey wrote:
>>> Rafael Vanoni Polanczyk wrote:
>>>
>>>  
>>>> Aubrey Li wrote:
>>>>    
>>>>> 2008/6/23 Rafael Vanoni <Rafael.Vanoni at sun.com>:
>>>>>      
>>>>>> Forgot to zero cstate_info[i].events, here's the correct diff.
>>>>>>
>>>>>>         
>>>>> Thanks for the great work, :-)
>>>>> This one looks good except this:
>>>>>
>>>>> -static char cpupm_treshold[]   = " echo cpu-threshold 1s >>
>>>>> /etc/power.conf"; +static char cpupm_treshold[]   = " echo
>>>>> cpu-threshold 5s >> /etc/power.conf";
>>>>> why do we need this change?
>>>>>
>>>>>       
>>>> Sorry I didn't comment this one before. 1s seems a bit too little,
>>>> don't you think? I can't remember the default, tho.
>>>>
>>>> thanks
>>>> rafael
>>>>     
>>> Not really, intel processor can switch very quickly (in microseconds).
>>> That's why we are working on the subsecond p-state. So here, we should
>>> set this threshold as small as possible, 1s is the right value.
>>
>> I've not really been following this thread. But why is this change 
>> being proposed and is the proposal that the change be a productized 
>> change to Solaris? If so, please don't. I'm currently looking into 
>> making this same change, enabling CPU power management by default and 
>> modifying the Solaris PM framework to scan CPU devices once a second:
>>
>> 6647538 cpupm should be on by default
>> 6714184 x86 CPU power management could be a little more aggressive.
>>
>> Just changing cpu-threshold to 1 second (as I think you are proposing 
>> above) isn't going to have much effect without the changes I'm 
>> planning with the CRs above. Also, I'm currently running these 
>> changes through our PerfPIT to see if they will have any impact on 
>> performance.
>
> Sounds good. We just have to pay some attention to this value in the 
> future. For instance, I have a white paper with me that suggested 
> setting the threshold to 15s.

I'm sure that the 15s is the whitepaper was taken from my blog. I 
suggested 15s in my blog because of the PM framework min scan of 15 seconds.

Mark


Reply via email to