Re: [PATCH v5 2/5] lib: Add zstd modules

[Austin S. Hemmelgarn]

On 2017-08-10 13:24, Eric Biggers wrote:
> On Thu, Aug 10, 2017 at 07:32:18AM -0400, Austin S. Hemmelgarn wrote:
> > On 2017-08-10 04:30, Eric Biggers wrote:
> > > On Wed, Aug 09, 2017 at 07:35:53PM -0700, Nick Terrell wrote:
> > > > It can compress at speeds approaching lz4, and quality approaching lzma.
> > >
> > > Well, for a very loose definition of "approaching", and certainly not at the
> > > same time.  I doubt there's a use case for using the highest compression levels
> > > in kernel mode --- especially the ones using zstd_opt.h.
> > Large data-sets with WORM access patterns and infrequent writes
> > immediately come to mind as a use case for the highest compression
> > level.
> >
> > As a more specific example, the company I work for has a very large
> > amount of documentation, and we keep all old versions.  This is all
> > stored on a file server which is currently using BTRFS.  Once a
> > document is written, it's almost never rewritten, so write
> > performance only matters for the first write.  However, they're read
> > back pretty frequently, so we need good read performance.  As of
> > right now, the system is set to use LZO compression by default, and
> > then when a new document is added, the previous version of that
> > document gets re-compressed using zlib compression, which actually
> > results in pretty significant space savings most of the time.  I
> > would absolutely love to use zstd compression with this system with
> > the highest compression level, because most people don't care how
> > long it takes to write the file out, but they do care how long it
> > takes to read a file (even if it's an older version).
>
> This may be a reasonable use case, but note this cannot just be the regular
> "zstd" compression setting, since filesystem compression by default must provide
> reasonable performance for many different access patterns.  See the patch in
> this series which actually adds zstd compression to btrfs; it only uses level 1.
> I do not see a patch which adds a higher compression mode.  It would need to be
> a special setting like "zstdhc" that users could opt-in to on specific
> directories.  It also would need to be compared to simply compressing in
> userspace.  In many cases compressing in userspace is probably the better
> solution for the use case in question because it works on any filesystem, allows
> using any compression algorithm, and if random access is not needed it is
> possible to compress each file as a single stream (like a .xz file), which
> produces a much better compression ratio than the block-by-block compression
> that filesystems have to use.
There has been discussion as well as (I think) initial patches merged 
for support of specifying the compression level for algorithms which 
support multiple compression levels in BTRFS.  I was actually under the 
impression that we had decided to use level 3 as the default for zstd, 
but that apparently isn't the case, and with the benchmark issues, it 
may not be once proper benchmarks are run.

Also, on the note of compressing in userspace, the use case I quoted at 
least can't do that because we have to deal with Windows clients and 
users have to be able to open files directly on said Windows clients.  I 
entirely agree that real archival storage is better off using userspace 
compression, but sometimes real archival storage isn't an option.
> Note also that LZ4HC is in the kernel source tree currently but no one is using
> it vs. the regular LZ4.  I think it is the kind of thing that sounded useful
> originally, but at the end of the day no one really wants to use it in kernel
> mode.  I'd certainly be interested in actual patches, though.
Part of that is the fact that BTRFS is one of the only consumers (AFAIK) 
of this API that can freely choose all aspects of their usage, and the 
consensus here (which I don't agree with I might add) amounts to the 
argument that 'we already have <X> compression with a <Y> compression 
ratio, we don't need more things like that'.  I would personally love to 
see LZ4HC support in BTRFS (based on testing my own use cases, LZ4 is 
more deterministic than LZO for both compression and decompression, and 
most of the non archival usage I have of BTRFS benefits from 
determinism), but there's not any point in me writing up such a patch 
because it's almost certain to get rejected because BTRFS already has 
LZO.  The main reason that zstd is getting considered at all is that the 
quoted benchmarks show clear benefits in decompression speed relative to 
zlib and far better compression ratios than LZO.