Re: [PATCH 3/3] KVM: Add capability to not exit on PAUSE

["Longpeng (Mike)" <longpeng2@huawei.com>]

On 2017/11/29 8:09, Jan H. Schönherr wrote:

> On 11/28/2017 04:37 AM, Longpeng (Mike) wrote:
>>
>> On 2017/11/25 21:09, Jan H. Schönherr wrote:
>>
>>> Allow to disable pause loop exit/pause filtering on a per VM basis.
>>>
>>> If some VMs have dedicated host CPUs, they won't be negatively affected
>>> due to needlessly intercepted PAUSE instructions.
>>>
>> Hi Jan,
>>
>> Is there any difference between 'disable PLE in vmcs' and 'make ple_gap per
>> VM/VCPU and set ple_gap=0 for vcpus which is dedicated' ?
> 
> "Just" disabling PLE in vmcs would still call into some of the PLE window
> adjustment paths and potentially do some VMCS writes at times. My patch should
> have eliminated these cases as well.
> 

Ah, I see, thanks. :)

We used a proprietary test suite to benchmark performance with ple_gap=0 and
ple_gap=1 (very little VMexits due to PLE in kvm trace), we found that ple_gap=1
is always better than ple_gap=0, we don't know why, maybe impacts the hardware
logical. Do you have any idea about this ?

> However, making all the PLE configuration knobs per VM has the difficulty
> that you'd need to be able to specify them in some way. That would not only
> be a x86 specific interface, but a VMX specific one as well. VMX-PLE and
> SVM-PF don't look compatible enough for a shared configuration.
> 
> That's why I only went for the binary on/off interface.
> 

Yes, I think your patch is pretty good.

> Regards
> Jan
> 
> 
> 
> .
> 


-- 
Regards,
Longpeng(Mike)