Re: Buffered IO async context overhead

[Jens Axboe <axboe@kernel.dk>]

On 3/9/20 3:02 PM, Pavel Begunkov wrote:
> On 09/03/2020 23:41, Jens Axboe wrote:
>> On 3/9/20 2:03 PM, Pavel Begunkov wrote:
>>> On 24/02/2020 18:22, Jens Axboe wrote:
>>> A problem here is that we actually have a 2D array of works because of linked
>>> requests.
>>
>> You could either skip anything with a link, or even just ignore it and
>> simply re-queue a dependent link if it isn't hashed when it's done if
>> grabbed in a batch.
>>
>>> We can io_wqe_enqueue() dependant works, if have hashed requests, so delegating
>>> it to other threads. But if the work->list is not per-core, it will hurt
>>> locality. Either re-enqueue hashed ones if there is a dependant work. Need to
>>> think how to do better.
>>
>> If we ignore links for a second, I think we can all agree that it'd be a
>> big win to do the batch.
> 
> Definitely
> 
>>
>> With links, worst case would then be something where every other link is
>> hashed.
>>
>> For a first patch, I'd be quite happy to just stop the batch if there's
>> a link on a request. The normal case here is buffered writes, and
>> that'll handle that case perfectly. Links will be no worse than before.
>> Seems like a no-brainer to me.
> 
> That isn't really a problem, just pointing that there could be
> optimisations for different cases.

Definitely, in case it wasn't clear, my suggestion was merely to go for
the main win and ignore the link side for now. It's best done separately
anyway, with the link stuff tweaking the batch behavior.

-- 
Jens Axboe