Re: [PATCH v2 0/5] fast poll multishot mode

[Jens Axboe]

On 5/6/22 8:33 PM, Jens Axboe wrote:
> 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.

Running some quick testing, on an actual test box (previous numbers were
from a vm on my laptop):

Testing singleshot, normal files
Did 10000 accepts

________________________________________________________
Executed in  216.10 millis    fish           external
   usr time    9.32 millis  150.00 micros    9.17 millis
   sys time  110.06 millis   67.00 micros  109.99 millis

Testing multishot, fixed files
Did 10000 accepts

________________________________________________________
Executed in  189.04 millis    fish           external
   usr time   11.86 millis  159.00 micros   11.71 millis
   sys time   93.71 millis   70.00 micros   93.64 millis

That's about ~19 usec to accept a connection, pretty decent. Using
singleshot and with fixed files, it shaves about ~8% off, ends at around
200msec.

I think we can get away with using fixed files and multishot, attaching
the quick patch I did below to test it. We need something better than
this, otherwise once the space fills up, we'll likely end up with a
sparse space and the naive approach of just incrementing the next slot
won't work at all.

-- 
Jens Axboe