On Fri, Jan 13 2017, Hannes Reinecke wrote:

On 01/13/2017 12:04 PM, Hannes Reinecke wrote:

quoted

On 01/13/2017 09:15 AM, Hannes Reinecke wrote:

quoted

On 01/11/2017 10:39 PM, Jens Axboe wrote:

quoted

Another year, another posting of this patchset. The previous posting
was here:

https://www.spinics.net/lists/kernel/msg2406106.html

(yes, I've skipped v5, it was fixes on top of v4, not the rework).

I've reworked bits of this to get rid of the shadow requests, thanks
to Bart for the inspiration. The missing piece, for me, was the fact
that we have the tags->rqs[] indirection array already. I've done this
somewhat differently, though, by having the internal scheduler tag
map be allocated/torn down when an IO scheduler is attached or
detached. This also means that when we run without a scheduler, we
don't have to do double tag allocations, it'll work like before.

The patchset applies on top of 4.10-rc3, or can be pulled here:

git://git.kernel.dk/linux-block blk-mq-sched.6

Well ... something's wrong here on my machine:

[ .. ]

Turns out that selecting CONFIG_DEFAULT_MQ_DEADLINE is the culprit;
switching to CONFIG_DEFAULT_MQ_NONE and selecting mq-deadline after
booting manually makes the problem go away.

So there is a race condition during device init and switching the I/O
scheduler.

But the results from using mq-deadline are promising; the performance
drop I've seen on older hardware seems to be resolved:

mq iosched:
 seq read : io=13383MB, bw=228349KB/s, iops=57087
 rand read : io=12876MB, bw=219709KB/s, iops=54927
 seq write: io=14532MB, bw=247987KB/s, iops=61996
 rand write: io=13779MB, bw=235127KB/s, iops=58781
mq default:
 seq read : io=13056MB, bw=222588KB/s, iops=55647
 rand read : io=12908MB, bw=220069KB/s, iops=55017
 seq write: io=13986MB, bw=238444KB/s, iops=59611
 rand write: io=13733MB, bw=234128KB/s, iops=58532
sq default:
 seq read : io=10240MB, bw=194787KB/s, iops=48696
 rand read : io=10240MB, bw=191374KB/s, iops=47843
 seq write: io=10240MB, bw=245333KB/s, iops=61333
 rand write: io=10240MB, bw=228239KB/s, iops=57059

measured on mpt2sas with SSD devices.

Perfect! Straight on the path of kill of non scsi-mq, then.

I'll fix up the async scan issue. The new mq schedulers don't really
behave differently in this regard, so I'm a bit puzzled. Hopefully it
reproduces here.

-- 
Jens Axboe