Re: [PATCH 0/2] kernfs: speed up concurrency performance

[Greg KH]

On Wed, Dec 02, 2020 at 10:58:35PM +0800, Fox Chen wrote:
> Hello,
> 
> kernfs is an important facillity to support pseudo file systems and cgroup. 
> Currently, with a global mutex, reading files concurrently from kernfs (e.g. /sys) 
> is very slow.
> 
> This problem is reported by Brice Goglin on thread:
> Re: [PATCH 1/4] drivers core: Introduce CPU type sysfs interface
> https://lore.kernel.org/lkml/X60dvJoT4fURcnsF@xxxxxxxxx/
> 
> I independently comfirmed this on a 96-core AWS c5.metal server.
> Do open+read+write on /sys/devices/system/cpu/cpu15/topology/core_id 1000 times.
> With a single thread it takes ~2.5 us for each open+read+close.
> With one thread per core, 96 threads running simultaneously takes 540 us 
> for each of the same operation (without much variation) -- 200x slower than the 
> single thread one. 
> 
> The problem can only be observed in large machines (>=16 cores).
> The more cores you have the slower it can be.
> 
> Perf shows that CPUs spend most of the time (>80%) waiting on mutex locks in 
> kernfs_iop_permission and kernfs_dop_revalidate.
> 
> This patchset contains the following 2 patches:
> 0001-kernfs-replace-the-mutex-in-kernfs_iop_permission-wi.patch
> 0002-kernfs-remove-mutex-in-kernfs_dop_revalidate.patch
> 
> 0001 replace the mutex lock in kernfs_iop_permission with a new rwlock and 
> 0002 removes the mutex lock in kernfs_dop_revalidate.
> 
> After applying this patchset, the multi-thread performance becomes linear with 
> the fastest one at ~30 us to the worst at ~150 us, very similar as I tested it
> on a normal ext4 file system with fastest one at ~20 us to slowest at ~100 us. 
> And I believe that is largely due to spin_locks in filesystems which are normal.
> 
> Although it's still slower than single thread, users can benefit from this 
> patchset, especially ones working on HPC realm with lots of cpu cores and want to
> fetch system information from sysfs.

Does this mean that the changes slow down the single-threaded case?  Or
that it's just not as good as the speed of a single-threaded access?

But anyway, thanks so much for looking into this, it should help the
crazy systems out today, which means the normal systems in 5 years will
really appreciate this :)

Some minor comments on the individual patches follow...

thanks,

greg k-h