Re: [PATCH] md/raid5: reduce lock contention in read_one_chunk()

[Gal Ofri <gal.ofri@storing.io>]

On Fri, 4 Jun 2021 15:43:48 +0800
Guoqing Jiang <jgq516@gmail.com> wrote:

> Just curious '8'  is chose for  group_thread_cnt. IIUC, group means one 
> numa node, and better to
> set it  to match the number of cores in one numa node in case better 
> performance is expected.
Since the worker threads has really low impact on 100%-read workloads,
group_thread_cnt has no effect here.  It performs just as good with
group_thread_cnt set to 0, 8 or 96.  Thinking about it, I should probably
omit that comment from the patch.

btw, in write workloads, the worker threads are actually contended by
device_lock, so choosing any number >8 often results in degraded
performance in my tests.  I have a p.o.c. branch that reduces the
contention by replacing device_lock with hash locks to remove that
contention, but it's a big-risky patch so I'd need to break it down first. 

Thanks, Gal