Re: [PATCH 2/2] block: create ioctl to discard-or-zeroout a range of blocks

From: Chris Mason
Date: Wed Mar 16 2016 - 18:24:52 EST

On Tue, Mar 15, 2016 at 05:51:17PM -0700, Chris Mason wrote:
> On Tue, Mar 15, 2016 at 07:30:14PM -0500, Eric Sandeen wrote:
> > On 3/15/16 7:06 PM, Linus Torvalds wrote:
> > > On Tue, Mar 15, 2016 at 4:52 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > >> >
> > >> > It is pretty clear that the onus is on the patch submitter to
> > >> > provide justification for inclusion, not for the reviewer/Maintainer
> > >> > to have to prove that the solution is unworkable.
> > > I agree, but quite frankly, performance is a good justification.
> > >
> > > So if Ted can give performance numbers, that's justification enough.
> > > We've certainly taken changes with less.
> >
> > I've been away from ext4 for a while, so I'm really not on top of the
> > mechanics of the underlying problem at the moment.
> >
> > But I would say that in addition to numbers showing that ext4 has trouble
> > with unwritten extent conversion, we should have an explanation of
> > why it can't be solved in a way that doesn't open up these concerns.
> >
> > XFS certainly has different mechanisms, but is the demonstrated workload
> > problematic on XFS (or btrfs) as well? If not, can ext4 adopt any of the
> > solutions that make the workload perform better on other filesystems?
> When I've benchmarked this in the past, doing small random buffered writes
> into an preallocated extent was dramatically (3x or more) slower on xfs
> than doing them into a fully written extent. That was two years ago,
> but I can redo it.

So I re-ran some benchmarks, with 4K O_DIRECT random ios on nvme (4.5
kernel). This is O_DIRECT without O_SYNC. I don't think xfs will do
commits for each IO into the prealloc file? O_SYNC makes it much
slower, so hopefully I've got this right.

The test runs for 60 seconds, and I used an iodepth of 4:

prealloc file: 32,000 iops
overwrite: 121,000 iops

If I bump the iodepth up to 512:

prealloc file: 33,000 iops
overwrite: 279,000 iops

For streaming writes, XFS converts prealloc to written much better when
the IO isn't random. You can start seeing the difference at 16K
sequential O_DIRECT writes, but really its not a huge impact. The worst
case is 4K:

prealloc file: 227MB/s
overwrite: 340MB/s

I can't think of sequential workloads where this will matter, since they
will either end up with bigger IO or the performance impact won't get