Re: RFC fuse waitq latency

[Bernd Schubert <bernd.schubert@fastmail.fm>]

[I removed the failing netapp/zufs CCs]

On 4/22/22 14:25, Miklos Szeredi wrote:
> On Mon, 28 Mar 2022 at 15:21, Bernd Schubert <bschubert@ddn.com> wrote:
>>
>> I would like to discuss the user thread wake up latency in
>> fuse_dev_do_read(). Profiling fuse shows there is room for improvement
>> regarding memory copies and splice. The basic profiling with flame graphs
>> didn't reveal, though, why fuse is so much
>> slower (with an overlay file system) than just accessing the underlying
>> file system directly and also didn't reveal why a single threaded fuse
>> uses less than 100% cpu, with the application on top of use also using
>> less than 100% cpu (simple bonnie++ runs with 1B files).
>> So I started to suspect the wait queues and indeed, keeping the thread
>> that reads the fuse device for work running for some time gives quite
>> some improvements.
> 
> Might be related: I experimented with wake_up_sync() that didn't meet
> my expectations.  See this thread:
> 
> https://lore.kernel.org/all/1638780405-38026-1-git-send-email-quic_pragalla@quicinc.com/#r
> 
> Possibly fuse needs some wake up tweaks due to its special scheduling
> requirements.

Thanks I will look at that as well. I have a patch with spinning and 
avoid of thread wake  that is almost complete and in my (still limited) 
testing almost does not take more CPU and improves meta data / bonnie 
performance in between factor ~1.9 and 3, depending on in which 
performance mode the cpu is.

https://github.com/aakefbs/linux/commits/v5.17-fuse-scheduling3

Missing is just another option for wake-queue-size trigger and handling 
of signals. Should be ready once I'm done with my other work.

That being said, in the mean time I do believe a better approach would 
be SQ/CQ like, similar to NVME or io_uring. In principle exactly as 
io_uring, just the other way around - kernel fills in SQ, user space 
consumes it and fills CQ. We also looked into zufs and your fuse2 branch 
and were almost ready to start to port it to a recent kernel, but it is 
still all systemcall based and has waitq's - probably much slower than 
what could be achieved through queue pairs. Assuming userspace would not 
want a polling thread, but would want a notification similar to 
io_uring_enter(), there would be still a thread needed to be woken up, 
may that is where wake_up_sync() would help.

Btw, the optional kernel polling thread in io_uring also has spinning...


Bernd