Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:18 as permitted sender) client-ip=2620:137:e000::1:18;
Message-ID: <f38291aa-bc6f-dafe-4765-0785b72e6c53@kernel.dk>
Date:   Fri, 6 May 2022 20:33:55 -0600
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux aarch64; rv:91.0) Gecko/20100101
 Thunderbird/91.8.1
Subject: Re: [PATCH v2 0/5] fast poll multishot mode
Content-Language: en-US
From:   Jens Axboe <axboe@kernel.dk>
To:     Hao Xu <haoxu.linux@gmail.com>, io-uring@vger.kernel.org
Cc:     Pavel Begunkov <asml.silence@gmail.com>,
        linux-kernel@vger.kernel.org
References: <20220506070102.26032-1-haoxu.linux@gmail.com>
 <9157fe69-b5d4-2478-7a0d-e037b5550168@kernel.dk>
 <08ff00da-b871-2f2a-7b23-c8b2621df9dd@kernel.dk>
In-Reply-To: <08ff00da-b871-2f2a-7b23-c8b2621df9dd@kernel.dk>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 5/6/22 5:26 PM, Jens Axboe wrote:
> On 5/6/22 4:23 PM, Jens Axboe wrote:
>> On 5/6/22 1:00 AM, Hao Xu wrote:
>>> Let multishot support multishot mode, currently only add accept as its
>>> first comsumer.
>>> theoretical analysis:
>>>   1) when connections come in fast
>>>     - singleshot:
>>>               add accept sqe(userpsace) --> accept inline
>>>                               ^                 |
>>>                               |-----------------|
>>>     - multishot:
>>>              add accept sqe(userspace) --> accept inline
>>>                                               ^     |
>>>                                               |--*--|
>>>
>>>     we do accept repeatedly in * place until get EAGAIN
>>>
>>>   2) when connections come in at a low pressure
>>>     similar thing like 1), we reduce a lot of userspace-kernel context
>>>     switch and useless vfs_poll()
>>>
>>>
>>> tests:
>>> Did some tests, which goes in this way:
>>>
>>>   server    client(multiple)
>>>   accept    connect
>>>   read      write
>>>   write     read
>>>   close     close
>>>
>>> Basically, raise up a number of clients(on same machine with server) to
>>> connect to the server, and then write some data to it, the server will
>>> write those data back to the client after it receives them, and then
>>> close the connection after write return. Then the client will read the
>>> data and then close the connection. Here I test 10000 clients connect
>>> one server, data size 128 bytes. And each client has a go routine for
>>> it, so they come to the server in short time.
>>> test 20 times before/after this patchset, time spent:(unit cycle, which
>>> is the return value of clock())
>>> before:
>>>   1930136+1940725+1907981+1947601+1923812+1928226+1911087+1905897+1941075
>>>   +1934374+1906614+1912504+1949110+1908790+1909951+1941672+1969525+1934984
>>>   +1934226+1914385)/20.0 = 1927633.75
>>> after:
>>>   1858905+1917104+1895455+1963963+1892706+1889208+1874175+1904753+1874112
>>>   +1874985+1882706+1884642+1864694+1906508+1916150+1924250+1869060+1889506
>>>   +1871324+1940803)/20.0 = 1894750.45
>>>
>>> (1927633.75 - 1894750.45) / 1927633.75 = 1.65%
>>>
>>>
>>> A liburing test is here:
>>> https://github.com/HowHsu/liburing/blob/multishot_accept/test/accept.c
>>
>> Wish I had seen that, I wrote my own! But maybe that's good, you tend to
>> find other issues through that.
>>
>> Anyway, works for me in testing, and I can see this being a nice win for
>> accept intensive workloads. I pushed a bunch of cleanup patches that
>> should just get folded in. Can you fold them into your patches and
>> address the other feedback, and post a v3? I pushed the test branch
>> here:
>>
>> https://git.kernel.dk/cgit/linux-block/log/?h=fastpoll-mshot
> 
> Quick benchmark here, accepting 10k connections:
> 
> Stock kernel
> real	0m0.728s
> user	0m0.009s
> sys	0m0.192s
> 
> Patched
> real	0m0.684s
> user	0m0.018s
> sys	0m0.102s
> 
> Looks like a nice win for a highly synthetic benchmark. Nothing
> scientific, was just curious.

One more thought on this - how is it supposed to work with
accept-direct? One idea would be to make it incrementally increasing.
But we need a good story for that, if it's exclusive to non-direct
files, then it's a lot less interesting as the latter is really nice win
for lots of files. If we can combine the two, even better.

-- 
Jens Axboe