Re: [PATCHSET v6 0/4] Split iowait into two states

[Pavel Begunkov <asml.silence@gmail.com>]

On 3/31/25 11:33, Christian Loehle wrote:
> On 3/31/25 10:02, Pavel Begunkov wrote:
>> On 8/19/24 16:39, Jens Axboe wrote:
>>> Hi,
>>>
>>> This is v6 of the patchset where the current in_iowait state is split
>>> into two parts:
>>>
>>> 1) The "task is sleeping waiting on IO", and would like cpufreq goodness
>>>      in terms of sleep and wakeup latencies.
>>> 2) The above, and also accounted as such in the iowait stats.
>>>
>>> The current ->in_iowait covers both, this series splits it into two types
>>> of state so that each can be controlled seperately.
>>>
>>> Patches 1..3 are prep patches, changing the type of
>>> task_struct->nr_iowait and adding helpers to manipulate the iowait counts.
>>>
>>> Patch 4 does the actual splitting.
>>>
>>> This has been sitting for a while, would be nice to get this queued up
>>> for 6.12. Comments welcome!
>>
>> Good day,
>>
>> Did anything good happened with these patches or related work?
>> Christian>
> 
> Hi Pavel,
> so for cpuidle part we've had commit ("38f83090f515 cpuidle: menu: Remove iowait influence")
> for a while now without much complaints, hopefully that means it stays in.
> So I'd really like to know how the results still compare for relevant workloads.

Sounds great

> cpufreq iowait boosting is still a thing in schedutil and intel_pstate,
> and so far I've failed to convince Rafael and Peter to get rid of it.
> I still think that is the right thing to do, but it does come with a
> regression in most of the simple synthetic fio tests.

IOW, from the io_uring iowait stat problem perspective it got stuck
and is unlikely to move short term.

>> Reminder: the goal is to let io_uring to keep using iowait boosting
>> but avoid reporting it in the iowait stats, because the jump in the
>> stat spooks users. I know at least several users carrying out of tree
>> patches to work it around. And, apparently, disabling the boosting
>> causes perf regressions.
> 
> Details would be appreciated, I looked the the postgres workload that
> justified it initially and that was on cpuidle iowait which is no
> longer a thing.

I wasn't involved and afraid don't have any extra numbers.

>> I'm reading through the thread, but unless I missed something, it looks
>> like the patchset is actually aligned with future plans on iowait
>> mentioned in the thread, in a sense that it reduces the exposure to
>> the user space, and, when it's time, a better approach will be able
>> replaces it with no visible effect to the user.
> 
> I'm not against $subject necessarily, it's clearly a hack tapering
> over this but as I've mentioned I'm fine carrying a revert of $subject
> for a future series on iowait boosting.
> 
>>
>> On the other hand, there seems to be a work around io_uring patch
>> queued for, which I quite dislike from io_uring perspective but also
>> because it exposes even more of iowait to the user.
>> I can understand why it's there, it has been over a year since v1,
>> but maybe we can figure something out before it's released? Would
>> it be fine to have something similar to this series? Any other
>> ideas?
> 
> Ah thank you, I've missed this
> https://lore.kernel.org/io-uring/f548f142-d6f3-46d8-9c58-6cf595c968fb@kernel.dk/
> Would be nice if this lead to more numbers comparing the two at least.

Sure, but I'd rather avoid adding this type of a uapi just to test
it and solve the problem a different way after.

-- 
Pavel Begunkov