Re: [PATCH] block: Call .initialize_rq_fn() also for filesystem requests

On Tue, Aug 29, 2017 at 08:15:23PM +0000, Bart Van Assche wrote:

On Tue, 2017-08-29 at 19:16 +0800, Ming Lei wrote:

quoted

Hi Bart,

Did you see perf regression on SRP with smaller jobs after applying
my patchset V3?

I just run the test with 16 jobs(the system has 16 CPU cores) instead of 64,
looks not see perf regression on SRP about v4.13-rc6+blk-next(2nd column) VS.
v4.13-rc6+blk-next+patch V3(3rd column):


---------------------------------------
 IOPS(K)  |    NONE     |    NONE
---------------------------------------
read      |      475.83 |      485.88
---------------------------------------
randread  |      142.86 |      141.96
---------------------------------------
write     |       483.9 |      492.39
---------------------------------------
randwrite |      124.83 |      124.53
---------------------------------------
[ ... ]
---------------------------------------
 LAT(us)  |    NONE     |    NONE
---------------------------------------
read      |        2.15 |        2.11
---------------------------------------
randread  |        7.17 |        7.21
---------------------------------------
write     |        2.11 |        2.08
---------------------------------------
randwrite |         8.2 |        8.22
---------------------------------------
[ ... ]

Hello Ming,

Although I would prefer to see measurement data against an SRP target system
that supports a higher workload (>1M IOPS) and

Hi Bart,

For so higher workload, I guess it often requires to increase
.cmd_per_lun.

also for a high-end NVMe drive,

My patch won't affect NVMe drive since NVMe driver doesn't become busy
usually.

I think the above data is sufficient to show that the performance impact of
your patch series is most likely small enough even for high-end SCSI initiator
drivers.

OK.

-- 
Ming