Re: [RFC 0/4] completion locking optimisation feature

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

On 3/18/22 15:22, Jens Axboe wrote:
> On 3/18/22 9:00 AM, Pavel Begunkov wrote:
>> On 3/18/22 14:52, Jens Axboe wrote:
>>> On 3/18/22 8:42 AM, Pavel Begunkov wrote:
>>>> On 3/18/22 13:52, Pavel Begunkov wrote:
>>>>> A WIP feature optimising out CQEs posting spinlocking for some use cases.
>>>>> For a more detailed description see 4/4.
>>>>>
>>>>> Quick benchmarking with fio/t/io_uring nops gives extra 4% to throughput for
>>>>> QD=1, and ~+2.5% for QD=4.
>>>>
>>>> Non-io_uring overhead (syscalls + userspace) takes ~60% of all execution
>>>> time, so the percentage should quite depend on the CPU and the kernel config.
>>>> Likely to be more than 4% for a faster setup.
>>>>
>>>> fwiw, was also usingIORING_ENTER_REGISTERED_RING, if it's not yet included
>>>> in the upstream version of the tool.
>>>
>>> But that seems to be exclusive of using PRIVATE_CQ?
>>
>> No, it's not. Let me ask to clarify the description and so, why do you
>> think it is exclusive?
> 
> Didn't dig too deep, just saw various EINVAL around registered files and
> updating/insertion. And the fact that t/io_uring gets -EINVAL on

Ah, those are for registered rsrc with tags, just because they use
fill_cqe, need to do something about them

> register with op 20 (ring fd) if I set CQ_PRIVATE, see other email.

-- 
Pavel Begunkov