On 10/04/2012 06:14 PM, Avi Kivity wrote:
> On 10/04/2012 12:56 PM, Raghavendra K T wrote:
>> On 10/03/2012 10:55 PM, Avi Kivity wrote:
>>> On 10/03/2012 04:29 PM, Raghavendra K T wrote:
>>>> * Avi Kivity <avi@redhat.com> [2012-09-27 14:03:59]:
>>>>
>>>>> On 09/27/2012 01:23 PM, Raghavendra K T wrote:
>>>>>>>
>>>> [...]
>>>>>> 2) looking at the result (comparing A & C) , I do feel we have
>>>>>> significant in iterating over vcpus (when compared to even vmexit)
>>>>>> so We still would need undercommit fix sugested by PeterZ
>>>>>> (improving by
>>>>>> 140%). ?
>>>>>
>>>>> Looking only at the current runqueue?  My worry is that it misses a lot
>>>>> of cases.  Maybe try the current runqueue first and then others.
>>>>>
>>>>
>>>> Okay. Do you mean we can have something like
>>>>
>>>> +       if (rq->nr_running == 1 && p_rq->nr_running == 1) {
>>>> +               yielded = -ESRCH;
>>>> +               goto out_irq;
>>>> +       }
>>>>
>>>> in the Peter's patch ?
>>>>
>>>> ( I thought lot about && or || . Both seem to have their own cons ).
>>>> But that should be only when we have short term imbalance, as PeterZ
>>>> told.
>>>
>>> I'm missing the context.  What is p_rq?
>>
>> p_rq is the run queue of target vcpu.
>> What I was trying below was to address Rik concern. Suppose
>> rq of source vcpu has one task, but target probably has two task,
>> with a eligible vcpu waiting to be scheduled.
>>
>>>
>>> What I mean was:
>>>
>>>     if can_yield_to_process_in_current_rq
>>>        do that
>>>     else if can_yield_to_process_in_other_rq
>>>        do that
>>>     else
>>>        return -ESRCH
>>
>> I think you are saying we have to check the run queue of the
>> source vcpu, if we have a vcpu belonging to same VM and try yield to
>> that? ignoring whatever the target vcpu we received for yield_to.
>>
>> Or is it that kvm_vcpu_yield_to should now check the vcpus of same vm
>> belonging to same run queue first. If we don't succeed, go again for
>> a vcpu in different runqueue.
>
> Right.  Prioritize vcpus that are cheap to yield to.  But may return bad
> results if all vcpus on the current runqueue are spinners, so probably
> not a good idea.

Okay. 'll drop vcpu from same rq idea now.

>
>> Does it add more overhead especially in <= 1x scenario?
>
> The current runqueue should have just our vcpu in that case, so low
> overhead.  But it's a bad idea due to the above scenario.
>