Re: [PATCH] io_uring: create percpu io sq thread when IORING_SETUP_SQ_AFF is flagged

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

On 22/05/2020 11:33, Xiaoguang Wang wrote:
> hi Pavel,
> 
> First thanks for reviewing!

sure, you're welcome!

>> On 20/05/2020 14:56, Xiaoguang Wang wrote:
>>>
>>> To fix this issue, when io_uring instance uses IORING_SETUP_SQ_AFF to specify a
>>> cpu,  we create a percpu io sq_thread to handle multiple io_uring instances' io
>>> requests serially. With this patch, in same environment, we get a huge
>>> improvement:
>>
>> Consider that a user can issue a long sequence of requests, say 2^15 of them,
>> and all of them will happen in a single call to io_submit_sqes(). Just preparing
>> them would take much time, apart from that it can do some effective work for
>> each of them, e.g. copying a page. And now there is another io_uring, caring
>> much about latencies and nevertheless waiting for _very_ long for its turn.
> Indeed I had thought this case before and don't clearly know how to optimize it
> yet.
> But I think if user has above usage scenarios, they probably shouldn't make
> io_uring
> instances share same cpu core. I'd like to explain more about our usage scenarios.

IIRC, you create _globally_ available per-cpu threads, so 2 totally independent
users may greatly affect each other, e.g. 2 docker containers in a hosted
server, and that would be nasty. But if it's limited to a "single user", it'd be
fine by me, see below for more details.

BTW, out of curiosity, what's the performance\latency impact of disabling SQPOLL
at all for your app? Say, comparing with non contended case.

> In a physical machine, say there are 96 cores, and it runs multiple cgroups, every
> cgroup run same application and will monopoly 16 cpu cores. This application will
> create 16 io threads and every io thread will create an io_uring instance and every
> thread will be bound to a different cpu core, these io threads will receive io
> requests.
> If we enable SQPOLL for these io threads, we allocate one or two cpu cores for
> these
> io_uring instances at most, so they must share allocated cpu core. It's totally
> disaster
> that some io_uring instances' busy loop in their sq_thread_idle period will
> impact other
> io_uring instances which have io requests to handle.
> 
>>
>> Another problem is that with it a user can't even guess when its SQ would be
>> emptied, and would need to constantly poll.
> In this patch, in every iteration, we only handle io requests already queued,
> will not constantly poll.

I was talking about a user polling _from userspace_ its full SQ, to understand
when it can submit more. That's if it doesn't want to wait for CQEs yet for some
reason (e.g. useful for net apps).

> 
>>
>> In essence, the problem is in bypassing thread scheduling, and re-implementing
>> poor man's version of it (round robin by io_uring).
> Yes :) Currently I use round robin strategy to handle multiple io_uring instance
> in every iteration.
> 
>> The idea and the reasoning are compelling, but I think we need to do something
>> about unrelated io_uring instances obstructing each other. At least not making
>> it mandatory behaviour.
>>
>> E.g. it's totally fine by me, if a sqpoll kthread is shared between specified
>> bunch of io_urings -- the user would be responsible for binding them and not
>> screwing up latencies/whatever. Most probably there won't be much (priviledged)
>> users using SQPOLL, and they all be a part of a single app, e.g. with
>> multiple/per-thread io_urings.
> Did you read my patch? In this patch, I have implemented this idea :)

Took a glance, may have overlooked things. I meant to do as in your patch, but
not sharing threads between ALL io_uring in the system, but rather between a
specified set of them. In other words, making yours @percpu_threads not global,
but rather binding to a set of io_urings.

e.g. create 2 independent per-cpu sets of threads. 1st one for {a1,a2,a3}, 2nd
for {b1,b2,b3}.

a1 = create_uring()
a2 = create_uring(shared_sq_threads=a1)
a3 = create_uring(shared_sq_threads=a1)

b1 = create_uring()
b2 = create_uring(shared_sq_threads=b1)
b3 = create_uring(shared_sq_threads=b1)

And then:
- it somehow solves the problem. As long as it doesn't effect much other users,
it's ok to let userspace screw itself by submitting 2^16 requests.

- there is still a problem with a simple round robin. E.g. >100 io_urings per
such set. Even though, a user may decide for itself, it worth to think about. I
don't want another scheduling framework here. E.g. first round-robin, then
weighted one, etc.

- it's actually not a set of threads (i.e. per-cpu) the API should operate on,
but just binding io_urings to a single SQPOLL thread.

- there is no need to restrict it to cpu-pinned case.

>>
>> Another way would be to switch between io_urings faster, e.g. after processing
>> not all requests but 1 or some N of them. But that's very thin ice, and I
>> already see other bag of issues.
> Sounds good, could you please lift detailed issues? Thanks.

Sounds terrible, TBH. Especially with arbitrary length links.


-- 
Pavel Begunkov