On 2/1/24 12:18 AM, Oliver Sang wrote:
> hi, Jens Axboe,
>
> On Wed, Jan 31, 2024 at 11:42:46AM -0700, Jens Axboe wrote:
>> On 1/31/24 11:17 AM, Bart Van Assche wrote:
>>> On 1/31/24 07:42, kernel test robot wrote:
>>>> kernel test robot noticed a -72.9% regression of fio.write_iops on:
>>>>
>>>>
>>>> commit: 574e7779cf583171acb5bf6365047bb0941b387c ("block/mq-deadline: use separate insertion lists")
>>>> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
>>>>
>>>> testcase: fio-basic
>>>> test machine: 64 threads 2 sockets Intel(R) Xeon(R) Gold 6346 CPU @ 3.10GHz (Ice Lake) with 256G memory
>>>> parameters:
>>>>
>>>> runtime: 300s
>>>> disk: 1HDD
>>>> fs: xfs
>>>> nr_task: 100%
>>>> test_size: 128G
>>>> rw: write
>>>> bs: 4k
>>>> ioengine: io_uring
>>>> direct: direct
>>>> cpufreq_governor: performance
>>>
>>> The actual test is available in this file:
>>> https://download.01.org/0day-ci/archive/20240131/202401312320.a335db14-oliver.sang@xxxxxxxxx/repro-script
>>>
>>> I haven't found anything in that file for disabling merging. Merging
>>> requests decreases IOPS. Does this perhaps mean that this test is
>>> broken?
>>
>> It's hard to know as nothing in this email or links include the actual
>> output of the job...
>
> I attached a dmesg and 2 outputs while running tests on 574e7779cf.
> not sure if they are helpful?
Both fio outputs is all I need, but I only see one of them attached?
>> But if it's fio IOPS, then those are application side and don't
>> necessarily correlate to drive IOPS due to merging. Eg for fio iops,
>> if it does 4k sequential and we merge to 128k, then the fio perceived
>> iops will be 32 times larger than the device side.
>>
>> I'll take a look, but seems like there might be something there. By
>> inserting into the other list, the request is also not available for
>> merging. And the test in question does single IOs at the time.
>
> if you have any debug patch want us to run, please just let us know.
> it will be our great pleasure!
Thanks, might take you up on that, probably won't have time for this
until next week however.
--
Jens Axboe
