[Drbd-dev] [PATCH v4 01/11] block: make generic_make_request handle arbitrarily sized bios

[Mike Snitzer]

On Tue, Jun 02 2015 at  4:59pm -0400,
Ming Lin <mlin at kernel.org> wrote:

> On Sun, May 31, 2015 at 11:02 PM, Ming Lin <mlin at kernel.org> wrote:
> > On Thu, 2015-05-28 at 01:36 +0100, Alasdair G Kergon wrote:
> >> On Wed, May 27, 2015 at 04:42:44PM -0700, Ming Lin wrote:
> >> > Here are fio results of XFS on a DM stripped target with 2 SSDs + 1 HDD.
> >> > Does it make sense?
> >>
> >> To stripe across devices with different characteristics?
> >>
> >> Some suggestions.
> >>
> >> Prepare 3 kernels.
> >>   O - Old kernel.
> >>   M - Old kernel with merge_bvec_fn disabled.
> >>   N - New kernel.
> >>
> >> You're trying to search for counter-examples to the hypothesis that
> >> "Kernel N always outperforms Kernel O".  Then if you find any, trying
> >> to show either that the performance impediment is small enough that
> >> it doesn't matter or that the cases are sufficiently rare or obscure
> >> that they may be ignored because of the greater benefits of N in much more
> >> common cases.
> >>
> >> (1) You're looking to set up configurations where kernel O performs noticeably
> >> better than M.  Then you're comparing the performance of O and N in those
> >> situations.
> >>
> >> (2) You're looking at other sensible configurations where O and M have
> >> similar performance, and comparing that with the performance of N.
> >
> > I didn't find case (1).
> >
> > But the important thing for this series is to simplify block layer
> > based on immutable biovecs. I don't expect performance improvement.

No simplifying isn't the important thing.  Any change to remove the
merge_bvec callbacks needs to not introduce performance regressions on
enterprise systems with large RAID arrays, etc.

It is fine if there isn't a performance improvement but I really don't
think the limited testing you've done on a relatively small storage
configuration has come even close to showing these changes don't
introduce performance regressions.

> > Here is the changes statistics.
> >
> > "68 files changed, 336 insertions(+), 1331 deletions(-)"
> >
> > I run below 3 test cases to make sure it didn't bring any regressions.
> > Test environment: 2 NVMe drives on 2 sockets server.
> > Each case run for 30 minutes.
> >
> > 2) btrfs radi0
> >
> > mkfs.btrfs -f -d raid0 /dev/nvme0n1 /dev/nvme1n1
> > mount /dev/nvme0n1 /mnt
> >
> > Then run 8K read.
> >
> > [global]
> > ioengine=libaio
> > iodepth=64
> > direct=1
> > runtime=1800
> > time_based
> > group_reporting
> > numjobs=4
> > rw=read
> >
> > [job1]
> > bs=8K
> > directory=/mnt
> > size=1G
> >
> > 2) ext4 on MD raid5
> >
> > mdadm --create /dev/md0 --level=5 --raid-devices=2 /dev/nvme0n1 /dev/nvme1n1
> > mkfs.ext4 /dev/md0
> > mount /dev/md0 /mnt
> >
> > fio script same as btrfs test
> >
> > 3) xfs on DM stripped target
> >
> > pvcreate /dev/nvme0n1 /dev/nvme1n1
> > vgcreate striped_vol_group /dev/nvme0n1 /dev/nvme1n1
> > lvcreate -i2 -I4 -L250G -nstriped_logical_volume striped_vol_group
> > mkfs.xfs -f /dev/striped_vol_group/striped_logical_volume
> > mount /dev/striped_vol_group/striped_logical_volume /mnt
> >
> > fio script same as btrfs test
> >
> > ------
> >
> > Results:
> >
> >         4.1-rc4         4.1-rc4-patched
> > btrfs   1818.6MB/s      1874.1MB/s
> > ext4    717307KB/s      714030KB/s
> > xfs     1396.6MB/s      1398.6MB/s
> 
> Hi Alasdair & Mike,
> 
> Would you like these numbers?
> I'd like to address your concerns to move forward.

I really don't see that these NVMe results prove much.

We need to test on large HW raid setups like a Netapp filer (or even
local SAS drives connected via some SAS controller).  Like a 8+2 drive
RAID6 or 8+1 RAID5 setup.  Testing with MD raid on JBOD setups with 8
devices is also useful.  It is larger RAID setups that will be more
sensitive to IO sizes being properly aligned on RAID stripe and/or chunk
size boundaries.

There are tradeoffs between creating a really large bio and creating a
properly sized bio from the start.  And yes, to one of neilb's original
points, limits do change and we suck at restacking limits.. so what was
once properly sized may no longer be but: that is a relatively rare
occurrence.  Late splitting does do away with the limits stacking
disconnect.  And in general I like the idea of removing all the
merge_bvec code.  I just don't think I can confidently Ack such a
wholesale switch at this point with such limited performance analysis.
If we (the DM/lvm team at Red Hat) are being painted into a corner of
having to provide our own testing that meets our definition of
"thorough" then we'll need time to carry out those tests.  But I'd hate
to hold up everyone because DM is not in agreement on this change...

So taking a step back, why can't we introduce late bio splitting in a
phased approach?

1: introduce late bio splitting to block core BUT still keep established
   merge_bvec infrastructure
2: establish a way for upper layers to skip merge_bvec if they'd like to
   do so (e.g. block-core exposes a 'use_late_bio_splitting' or
   something for userspace or upper layers to set, can also have a
   Kconfig that enables this feature by default)
3: we gain confidence in late bio-splitting and then carry on with the
   removal of merge_bvec et al (could be incrementally done on a
   per-driver basis, e.g. DM, MD, btrfs, etc, etc).

Mike