[bug-report] 5-9% FIO randomwrite ext4 perf regression on 6.12.y kernel

[Jens Axboe]

On 11/21/24 4:30 AM, Phil Auld wrote:
> 
> Hi,
> 
> On Wed, Nov 20, 2024 at 06:20:12PM -0700 Jens Axboe wrote:
>> On 11/20/24 5:00 PM, Chaitanya Kulkarni wrote:
>>> On 11/20/24 13:35, Saeed Mirzamohammadi wrote:
>>>> Hi,
>>>>
>>>> I?m reporting a performance regression of up to 9-10% with FIO randomwrite benchmark on ext4 comparing 6.12.0-rc2 kernel and v5.15.161. Also, standard deviation after this change grows up to 5-6%.
>>>>
>>>> Bisect root cause commit
>>>> ===================
>>>> - commit 63dfa1004322 ("nvme: move NVME_QUIRK_DEALLOCATE_ZEROES out of nvme_config_discard?)
>>>>
>>>>
>>>> Test details
>>>> =========
>>>> - readwrite=randwrite bs=4k size=1G ioengine=libaio iodepth=16 direct=1 time_based=1 ramp_time=180 runtime=1800 randrepeat=1 gtod_reduce=1
>>>> - Test is on ext4 filesystem
>>>> - System has 4 NVMe disks
>>>>
>>>
>>> Thanks a lot for the report, to narrow down this problem can you
>>> please :-
>>>
>>> 1. Run the same test on the raw nvme device /dev/nvme0n1 that you
>>>     have used for this benchmark ?
>>> 2. Run the same test on the  XFS formatted nvme device instead of ext4 ?
>>>
>>> This way we will know if there is an issue only with the ext4 or
>>> with other file systems are suffering from this problem too or
>>> it is below the file system layer such as block layer and nvme pci driver ?
>>>
>>> It will also help if you can repeat these numbers for io_uring fio io_engine
>>> to narrow down this problem to know if the issue is ioengine specific.
>>>
>>> Looking at the commit [1], it only sets the max value to write zeroes 
>>> sectors
>>> if NVME_QUIRK_DEALLOCATE_ZEROES is set, else uses the controller max
>>> write zeroes value.
>>
>> There's no way that commit is involved, the test as quoted doesn't even
>> touch write zeroes. Hence if there really is a regression here, then
>> it's either not easily bisectable, some error was injected while
>> bisecting, or the test itself is bimodal.
> 
> I was just going to ask how confident we are in that bisect result.
> 
> I suspect this is the same issue I've been fighting here:
> 
> https://lore.kernel.org/lkml/20241101124715.GA689589@pauld.westford.csb/
> 
> Saeed, can you try your randwrite test after
> 
>   "echo NO_DELAY_DEQUEUE > /sys/kernel/debug/sched/features"
> 
> please?
> 
> We don't as yet have a general fix for it as it seems to be a bit of
> a trade off.

Interesting. Might explain some regressions I've seen too related to
performance.

-- 
Jens Axboe