Re: [PATCH 2/3] blk-mq: Always complete remote completions requests in softirq

On 12/17/20 9:49 AM, Daniel Wagner wrote:

On Thu, Dec 17, 2020 at 09:45:47AM -0700, Jens Axboe wrote:

quoted

On 12/8/20 1:44 AM, Daniel Wagner wrote:

quoted

It looks like the patched version show tiny bit better numbers for this
workload. slat seems to be one of the major difference between the two
runs. But that is the only thing I really spotted to be really off.

slat is the same, just one is in nsec and the other is in usec.

Ah, good eyes. Need to remember this :)

quoted

quoted

I keep going with some more testing. Let what kind of tests you would
also want to see. I'll do a few plain NVMe tests next.

This is a good test, thanks.

Got sidetracked. Haven't started yet with these tests.

I just ran some here, and as expected, didn't see much change. Single
core IOPS at 2.8M in both cases, and not expecting any remote IPI for
that test case.

So I'd say that's good enough, wasn't expecting a change, and we don't
see one even for your case with remote IPI.

-- 
Jens Axboe