Re: [PATCH 1/4] staging: lustre: obdclass: change spinlock of key to rwlock

[NeilBrown <neilb@suse.com>]

[-- Attachment #1: Type: text/plain, Size: 1769 bytes --]

On Fri, May 04 2018, Dilger, Andreas wrote:

> On May 3, 2018, at 07:50, David Laight <David.Laight@ACULAB.COM> wrote:
>> 
>> From: James Simmons
>>> Sent: 02 May 2018 19:22
>>> From: Li Xi <lixi@ddn.com>
>>> 
>>> Most of the time, keys are never changed. So rwlock might be
>>> better for the concurrency of key read.
>> 
>> OTOH unless there is contention on the spin lock during reads the
>> additional cost of a rwlock (probably double that of a spinlock)
>> will hurt performance.
>> 
>> ...
>>> -	spin_lock(&lu_keys_guard);
>>> +	read_lock(&lu_keys_guard);
>>> 	atomic_inc(&lu_key_initing_cnt);
>>> -	spin_unlock(&lu_keys_guard);
>>> +	read_unlock(&lu_keys_guard);
>> 
>> WTF, seems unlikely that you need to hold any kind of lock
>> over an atomic_inc().
>> 
>> If this is just ensuring that no code holds the lock then
>> it would need to request the write_lock().
>> (and would need a comment)
>
> There was a fair amount of benchmarking done for this that shows the
> performance is significantly improved with the patch, which can be
> seen in the ticket that was referenced in the original commit comment:
>
> https://jira.hpdd.intel.com/browse/LU-6800?focusedCommentId=121776#comment-121776

That does surprise me.  The only places where the lock is held for read
are very short - clearing a few fields or incrementing a value.
But numbers don't lie.
I wonder if the next patch would have had just as big an effect. Taking
and dropping the lock 40 times is not likely to be good for performance.

Thanks,
NeilBrown


>
> That said, it might be good to include this information into the
> commit comment itself.
>
> Cheers, Andreas
> --
> Andreas Dilger
> Lustre Principal Architect
> Intel Corporation

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]