Re: [linus:master] [block] e70c301fae: stress-ng.aiol.ops_per_sec 49.6% regression

[Oliver Sang <oliver.sang@intel.com>]

hi, Niklas,

On Fri, Jan 03, 2025 at 10:09:14AM +0100, Niklas Cassel wrote:
> On Fri, Jan 03, 2025 at 07:49:25AM +0100, Christoph Hellwig wrote:
> > On Thu, Jan 02, 2025 at 10:49:41AM +0100, Niklas Cassel wrote:
> > > > > from below information, it seems an 'ahci' to me. but since I have limited
> > > > > knowledge about storage driver, maybe I'm wrong. if you want more information,
> > > > > please let us know. thanks a lot!
> > > > 
> > > > Yes, this looks like ahci.  Thanks a lot!
> > > 
> > > Did this ever get resolved?
> > > 
> > > I haven't seen a patch that seems to address this.
> > > 
> > > AHCI (ata_scsi_queuecmd()) only issues a single command, so if there is any
> > > reordering when issuing a batch of commands, my guess is that the problem
> > > also affects SCSI / the problem is in upper layers above AHCI, i.e. SCSI lib
> > > or block layer.
> > 
> > I started looking into this before the holidays.  blktrace shows perfectly
> > sequential writes without any reordering using ahci, directly on the
> > block device or using xfs and btrfs when using dd.  I also started
> > looking into what the test does and got as far as checking out the
> > stress-ng source tree and looking at stress-aiol.c.  AFAICS the default
> > submission does simple reads and writes using increasing offsets.
> > So if the test result isn't a fluke either the aio code does some
> > weird reordering or btrfs does.
> > 
> > Oliver, did the test also show any interesting results on non-btrfs
> > setups?
> > 
> 
> One thing that came to mind.
> Some distros (e.g. Fedora and openSUSE) ship with an udev rule that sets
> the I/O scheduler to BFQ for single-queue HDDs.
> 
> It could very well be the I/O scheduler that reorders.
> 
> Oliver, which I/O scheduler are you using?
> $ cat /sys/block/sdb/queue/scheduler 
> none mq-deadline kyber [bfq]

while our test running:

# cat /sys/block/sdb/queue/scheduler
none [mq-deadline] kyber bfq

> 
> 
> Kind regards,
> Niklas