On 2018/11/12 23:46, Dave Hansen wrote:
> On 11/11/18 9:38 PM, Li, Aubrey wrote:
>
>>> Do we want this, or do we want something more time-based?
>>>
>> This counter is introduced here to solve the race of context switch and
>> VZEROUPPER. 3 context switches mean the same thread is on-off CPU 3 times.
>> Due to scheduling latency, 3 jiffies could only happen AVX task on-off just
>> 1 time. So IMHO the context switches number is better here.
>
> Imagine we have a HZ=1000 system where AVX_STATE_DECAY_COUNT=3. That
> means that a task can be marked as a non-AVX-512-user after not using it
> for ~3 ms. But, with HZ=250, that's ~12ms.
>From the other side, if we set a 4ms decay, when HZ=1000, context switch count
is 4, that means, we have 4 times of chance to maintain the AVX state, that is,
we are able to filter 4 times init state reset out. But if HZ = 250, the context
switch is 1, we only have 1 time of chance to filter init state reset out.
>
> Also, don't forget that we have context switches from the timer
> interrupt, but also from normal old operations that sleep.
>
> Let's say our AVX-512 app was doing:
>
> while (foo) {
> do_avx_512();
> read(pipe, buf, len);
> read(pipe, buf, len);
> read(pipe, buf, len);
> }
>
> And all three pipe reads context-switched the task. That loop could
> finish in way under 3HZ, but still end up in do_avx_512() each time with
> fpu...avx->state=0.
Yeah, we are trying to address a prediction according to the historical pattern,
so you always can make a pattern to beat the prediction pattern. But in
practice,
I measured tensorflow with AVX512 enabled, linpack with AVX512, and a micro
benchmark, the current 3 context switches decay works well enough.
>
> BTW, I don't have a great solution for this. I was just pointing out
> one of the pitfalls from using context switch counts so strictly.
I really don't think time-based is better than the count in this case.
>>>> +/*
>>>> * Highest level per task FPU state data structure that
>>>> * contains the FPU register state plus various FPU
>>>> * state fields:
>>>> @@ -303,6 +312,14 @@ struct fpu {
>>>> unsigned char initialized;
>>>>
>>>> /*
>>>> + * @avx_state:
>>>> + *
>>>> + * This data structure indicates whether this context
>>>> + * contains AVX states
>>>> + */
>>>
>>> Yeah, that's precisely what fpu->state.xsave.xfeatures does. :)
>>> I see, will refine in the next version
>
> One other thought about the new 'avx_state':
>
> fxregs_state (which is a part of the XSAVE state) has some padding and
> 'sw_reserved' areas. You *might* be able to steal some space there.
> Not that this is a huge space eater, but why waste the space if we don't
> have to?
>
IMHO, I prefer not adding any extra thing into a data structure associated
with a hardware table. Let me try to work out a new version to see if it can
satisfy you.
Thanks,
-Aubrey
