[Intel-gfx] [PATCH i-g-t 4/8] tests/gem_close_race: Tune down for BAT ~1s.

[Mika Kuoppala]

Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com> writes:

> [ text/plain ]
>
> On 27/05/16 13:16, Chris Wilson wrote:
>> On Fri, May 27, 2016 at 01:10:07PM +0100, Tvrtko Ursulin wrote:
>>>
>>> On 27/05/16 12:58, Chris Wilson wrote:
>>>> On Fri, May 27, 2016 at 02:50:34PM +0300, Marius Vlad wrote:
>>>>> Signed-off-by: Marius Vlad <marius.c.vlad at intel.com>
>>>>
>>>> Nak. It's a race detector. Please suggest how to increase detection
>>>> rates.
>>>
>>> As a more or less well know TV personality would say - "it's better
>>> than nothing"! :))
>>>
>>> Seriously, under the new rules this is all we can do. Full
>>> gem-close-race will (will it?) run in the nightly run so a little
>>> bit more pain while bisecting breakages but those are the rules. We
>>> can give them a go and see how it works out.
>>
>> Wrong approach. Right approach would be to add a new test that reliably
>> detected a checklist of the most common races in under 1s. Nerfing a
>> test to make it useless makes BAT equally useless.
>>
>> So we are giving up on BAT?
>
> If it is not possible to stuff everything into the allocated budget, and 
> in cases where it may not be possible to come up with a 1s race 
> detector, it makes sense to move the test out of BAT and into the 
> nightly runs.
>

With tests to poke out races, I think the best way is to take
them out from BAT and move to nightly.

As A is for acceptance, it is not much value as some poor soul
will hit the race 1 month after we regressed it.

- Mika


> Agreement was that the time limits are hard limits so thats pretty much 
> it. We can do this now and hit the targets and them add more tests if 
> and when someone manages to implement them.
>
> I don't have a problem with that, B in BAT stands for basic anyway.
>
> We just adjust the model that when nightly test run fails someone from 
> QA gets tasked with bisection etc as a top priority.
>
> Regards,
>
> Tvrtko