Re: [PATCH] block/mq-deadline: Optimize request insertion

[Jens Axboe <axboe@kernel.dk>]

On 1/22/24 7:30 PM, Damien Le Moal wrote:
> On 1/23/24 08:53, Bart Van Assche wrote:
>> Reduce lock contention on dd->lock by calling dd_insert_request() from
>> inside the dispatch callback instead of from the insert callback. This
>> patch is inspired by a patch from Jens.
> 
> I supposed this is a followup of the performance discussion with Jens. If so,
> can you add performance numbers here and so justifying the change ?
> Otherwise, it is hard to figure out the effect of the patch and so why you are
> making this change.

Numbers are in here [1], I will fold them into the commit message. It's worth,
as is mentioned in my commit message, that this also relies on the previous
patches to get there. It just leaves the insertion side contention as the
dominant one until this patch is applied.

Before:

Device		IOPS	sys	contention	diff
====================================================
null_blk	879K	89%	93.6%
nvme0n1		901K	86%	94.5%

and after this and the previous two patches:

Device		IOPS	sys	contention	diff
====================================================
null_blk	2867K	11.1%	~6.0%		+226%
nvme0n1		3162K	 9.9%	~5.0%		+250%

[1] https://git.kernel.dk/cgit/linux/commit/?h=block-deadline&id=678abafc92bae49e6f4a92cea05dcd5f39928054

-- 
Jens Axboe