On 7/18/2017 10:36 AM, Keith Busch wrote:

On Mon, Jul 17, 2017 at 07:07:00PM -0400, okaya at codeaurora.org wrote:

quoted

Maybe, I need to understand the design better. I was curious why completion
and submission queues were protected by a single lock causing lock
contention.

Ideally the queues are tied to CPUs, so you couldn't have one thread
submitting to a particular queue-pair while another thread is reaping
completions from it. Such a setup wouldn't get lock contention.

I do see that the NVMe driver is creating a completion interrupt on
each CPU core for the completions. No problems with that. 

However, I don't think you can guarantee that there will always be a single
CPU core targeting one submission queue especially with asynchronous IO.

Lock contention counters from CONFIG_LOCK_STAT are pointing to nvmeq->lock
in my FIO tests.

Did I miss something?

Some machines have so many CPUs, though, that sharing hardware queues
is required. We've experimented with separate submission and completion
locks for such cases, but I've never seen an improved performance as a
result.

I have also experimented with multiple locks with no significant gains. 
However, I was curious if somebody else had a better implementation than mine.

-- 
Sinan Kaya
Qualcomm Datacenter Technologies, Inc. as an affiliate of Qualcomm Technologies, Inc.
Qualcomm Technologies, Inc. is a member of the Code Aurora Forum, a Linux Foundation Collaborative Project.