Re: [PATCH v1] io_uring: Add support for napi_busy_poll

[Jens Axboe <axboe@kernel.dk>]

On 2/20/22 11:37 AM, Olivier Langlois wrote:
> On Sat, 2022-02-19 at 17:22 -0700, Jens Axboe wrote:
>>
>> Outside of this, I was hoping to see some performance numbers in the
>> main patch. Sounds like you have them, can you share?
>>
> Yes.
> 
> It is not much. Only numbers from my application and it is far from
> being the best benchmark because the result can be influenced by
> multiple external factors.
> 
> Beside addressing the race condition remaining inside io_cqring_wait()
> around napi_list for v2 patch, creating a benchmark program that
> isolate the performance of the new feature is on my todo list.
> 
> I would think that creating a simple UDP ping-pong setup and measure
> RTT with and without busy_polling should be a good enough test.

Yes, a separate targeted test like that would be very useful and
interesting indeed!

> In the meantime, here are the results that I have:
> 
> Without io_uring busy poll:
> reaction time to an update: 17159usec
> reaction time to an update: 19068usec
> reaction time to an update: 23055usec
> reaction time to an update: 16511usec
> reaction time to an update: 17604usec
> 
> With io_uring busy poll:
> reaction time to an update: 15782usec
> reaction time to an update: 15337usec
> reaction time to an update: 15379usec
> reaction time to an update: 15275usec
> reaction time to an update: 15107usec

OK, that's a pretty good improvement in both latency and
deviation/consistency. Is this using SQPOLL, or is it using polling off
cqring_wait from the task itself? Also something to consider for the
test benchmark app, should be able to run both (which is usually just
setting the SETUP_SQPOLL flag or not, if done right).

> Concerning my latency issue with busy polling, I have found this that
> might help me:
> https://lwn.net/ml/netdev/20201002222514.1159492-1-weiwan@google.com/
> 

-- 
Jens Axboe