Re: [PATCH RFC v2 3/3] io_uring: batch get(ctx->ref) across submits

From: Pavel Begunkov <asml.silence@gmail.com>
To: Jens Axboe <axboe@kernel.dk>,
	io-uring@vger.kernel.org, linux-kernel@vger.kernel.org
Cc: Tejun Heo <tj@kernel.org>, Dennis Zhou <dennis@kernel.org>,
	Christoph Lameter <cl@linux.com>
Subject: Re: [PATCH RFC v2 3/3] io_uring: batch get(ctx->ref) across submits
Date: Sat, 21 Dec 2019 20:26:22 +0300	[thread overview]
Message-ID: <b5c0e2ab-ded3-d06b-afda-e7a72f1368e4@gmail.com> (raw)
In-Reply-To: <fef3a245-d2a2-23b3-ff03-3e05af19b752@kernel.dk>

[-- Attachment #1.1: Type: text/plain, Size: 2970 bytes --]

On 21/12/2019 20:01, Jens Axboe wrote:
> On 12/21/19 9:48 AM, Pavel Begunkov wrote:
>> On 21/12/2019 19:38, Jens Axboe wrote:
>>> On 12/21/19 9:20 AM, Pavel Begunkov wrote:
>>>> On 21/12/2019 19:15, Pavel Begunkov wrote:
>>>>> Double account ctx->refs keeping number of taken refs in ctx. As
>>>>> io_uring gets per-request ctx->refs during submission, while holding
>>>>> ctx->uring_lock, this allows in most of the time to bypass
>>>>> percpu_ref_get*() and its overhead.
>>>>
>>>> Jens, could you please benchmark with this one? Especially for offloaded QD1
>>>> case. I haven't got any difference for nops test and don't have a decent SSD
>>>> at hands to test it myself. We could drop it, if there is no benefit.
>>>>
>>>> This rewrites that @extra_refs from the second one, so I left it for now.
>>>
>>> Sure, let me run a peak test, qd1 test, qd1+sqpoll test on
>>> for-5.6/io_uring, same branch with 1-2, and same branch with 1-3. That
>>> should give us a good comparison. One core used for all, and we're going
>>> to be core speed bound for the performance in all cases on this setup.
>>> So it'll be a good comparison.
>>>
>> Great, thanks!
> 
> For some reason, not seeing much of a change between for-5.6/io_uring
> and 1+2 and 1+2+3, it's about the same and results seem very stable.
> For reference, top of profile with 1-3 applied looks like this:

I see. I'll probably drop the last one, as it only complicates things.

My apologies for misleading terminology. Read-only QD1 (submit and
wait until the userspace completes it) obviously won't saturate a CPU.
Writes probably wouldn't as well (though, depends on HW). And it would be
better to say -- submit by one, complete in a bunch.
Just curious, what you used for testing? Is it fio?

> 
> +    3.92%  io_uring  [kernel.vmlinux]  [k] blkdev_direct_IO
> +    3.87%  io_uring  [kernel.vmlinux]  [k] blk_mq_get_request
> +    3.43%  io_uring  [kernel.vmlinux]  [k] io_iopoll_getevents
> +    3.03%  io_uring  [kernel.vmlinux]  [k] __slab_free
> +    2.87%  io_uring  io_uring          [.] submitter_fn
> +    2.79%  io_uring  [kernel.vmlinux]  [k] io_submit_sqes
> +    2.75%  io_uring  [kernel.vmlinux]  [k] bio_alloc_bioset
> +    2.70%  io_uring  [nvme_core]       [k] nvme_setup_cmd
> +    2.59%  io_uring  [kernel.vmlinux]  [k] blk_mq_make_request
> +    2.46%  io_uring  [kernel.vmlinux]  [k] io_prep_rw
> +    2.32%  io_uring  [kernel.vmlinux]  [k] io_read
> +    2.25%  io_uring  [kernel.vmlinux]  [k] blk_mq_free_request
> +    2.19%  io_uring  [kernel.vmlinux]  [k] io_put_req
> +    2.06%  io_uring  [kernel.vmlinux]  [k] kmem_cache_alloc
> +    2.01%  io_uring  [kernel.vmlinux]  [k] generic_make_request_checks
> +    1.90%  io_uring  [kernel.vmlinux]  [k] __sbitmap_get_word
> +    1.86%  io_uring  [kernel.vmlinux]  [k] sbitmap_queue_clear
> +    1.85%  io_uring  [kernel.vmlinux]  [k] io_issue_sqe
> 
> 

-- 
Pavel Begunkov

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]