dwarfs

A fast high compression read-only file system

mhx
554
20
C++

Latest Release
Total Downloads
DwarFS CI Build
Build Status
Codacy Badge
codecov
OpenSSF Best Practices

DwarFS

The Deduplicating Warp-speed Advanced Read-only File System.

A fast high compression read-only file system for Linux and Windows.

Table of contents

Overview

Windows Screen Capture

Linux Screen Capture

DwarFS is a read-only file system with a focus on achieving very
high compression ratios in particular for very redundant data.

This probably doesn’t sound very exciting, because if it’s redundant,
it should compress well. However, I found that other read-only,
compressed file systems don’t do a very good job at making use of
this redundancy. See here for a comparison with other
compressed file systems.

DwarFS also doesn’t compromise on speed and for my use cases I’ve
found it to be on par with or perform better than SquashFS. For my
primary use case, DwarFS compression is an order of magnitude better
than SquashFS compression, it’s 6 times faster to build the file
system, it’s typically faster to access files on DwarFS and it uses
less CPU resources.

To give you an idea of what DwarFS is capable of, here’s a quick comparison
of DwarFS and SquashFS on a set of video files with a total size of 39 GiB.
The twist is that each unique video file has two sibling files with a
different set of audio streams (this is
an actual use case).
So there’s redundancy in both the video and audio data, but as the streams
are interleaved and identical blocks are typically very far apart, it’s
challenging to make use of that redundancy for compression. SquashFS
essentially fails to compress the source data at all, whereas DwarFS is
able to reduce the size by almost a factor of 3, which is close to the
theoretical maximum:

$ du -hs dwarfs-video-test
39G     dwarfs-video-test
$ ls -lh dwarfs-video-test.*fs
-rw-r--r-- 1 mhx users 14G Jul  2 13:01 dwarfs-video-test.dwarfs
-rw-r--r-- 1 mhx users 39G Jul 12 09:41 dwarfs-video-test.squashfs

Furthermore, when mounting the SquashFS image and performing a random-read
throughput test using fio-3.34, both
squashfuse and squashfuse_ll top out at around 230 MiB/s:

$ fio --readonly --rw=randread --name=randread --bs=64k --direct=1 \
      --opendir=mnt --numjobs=4 --ioengine=libaio --iodepth=32 \
      --group_reporting --runtime=60 --time_based
[...]
   READ: bw=230MiB/s (241MB/s), 230MiB/s-230MiB/s (241MB/s-241MB/s), io=13.5GiB (14.5GB), run=60004-60004msec

In comparison, DwarFS manages to sustain random read rates of 20 GiB/s:

  READ: bw=20.2GiB/s (21.7GB/s), 20.2GiB/s-20.2GiB/s (21.7GB/s-21.7GB/s), io=1212GiB (1301GB), run=60001-60001msec

Distinct features of DwarFS are:

  • Clustering of files by similarity using a similarity hash function.
    This makes it easier to exploit the redundancy across file boundaries.

  • Segmentation analysis across file system blocks in order to reduce
    the size of the uncompressed file system. This saves memory when
    using the compressed file system and thus potentially allows for
    higher cache hit rates as more data can be kept in the cache.

  • Categorization framework to categorize
    files or even fragments of files and then process individual categories
    differently. For example, this allows you to not waste time trying to
    compress incompressible files or to compress PCM audio data using FLAC
    compression.

  • Highly multi-threaded implementation. Both the
    file system creation tool as well as the
    FUSE driver are able to make good use of the
    many cores of your system.

History

I started working on DwarFS in 2013 and my main use case and major
motivation was that I had several hundred different versions of Perl
that were taking up something around 30 gigabytes of disk space, and
I was unwilling to spend more than 10% of my hard drive keeping them
around for when I happened to need them.

Up until then, I had been using Cromfs
for squeezing them into a manageable size. However, I was getting more
and more annoyed by the time it took to build the filesystem image
and, to make things worse, more often than not it was crashing after
about an hour or so.

I had obviously also looked into SquashFS,
but never got anywhere close to the compression rates of Cromfs.

This alone wouldn’t have been enough to get me into writing DwarFS,
but at around the same time, I was pretty obsessed with the recent
developments and features of newer C++ standards and really wanted
a C++ hobby project to work on. Also, I’ve wanted to do something
with FUSE
for quite some time. Last but not least, I had been thinking about
the problem of compressed file systems for a bit and had some ideas
that I definitely wanted to try.

The majority of the code was written in 2013, then I did a couple
of cleanups, bugfixes and refactors every once in a while, but I
never really got it to a state where I would feel happy releasing
it. It was too awkward to build with its dependency on Facebook’s
(quite awesome) folly library
and it didn’t have any documentation.

Digging out the project again this year, things didn’t look as grim
as they used to. Folly now builds with CMake and so I just pulled
it in as a submodule. Most other dependencies can be satisfied
from packages that should be widely available. And I’ve written
some rudimentary docs as well.

Building and Installing

Prebuilt Binaries

Each release has pre-built,
statically linked binaries for Linux-x86_64, Linux-aarch64 and
Windows-AMD64 available for download. These should run without
any dependencies and can be useful especially on older distributions
where you can’t easily build the tools from source.

Universal Binaries

In addition to the binary tarballs, there’s a universal binary
available for each architecture. These universal binaries contain
all tools (mkdwarfs, dwarfsck, dwarfsextract and the dwarfs
FUSE driver) in a single executable. These executables are compressed
using upx, so they are much smaller than
the individual tools combined. However, it also means the binaries need
to be decompressed each time they are run, which can have a signficant
overhead. If that is an issue, you can either stick to the “classic”
individual binaries or you can decompress the universal binary, e.g.:

upx -d dwarfs-universal-0.7.0-Linux-aarch64

The universal binaries can be run through symbolic links named after
the proper tool. e.g.:

$ ln -s dwarfs-universal-0.7.0-Linux-aarch64 mkdwarfs
$ ./mkdwarfs --help

This also works on Windows if the file system supports symbolic links:

> mklink mkdwarfs.exe dwarfs-universal-0.7.0-Windows-AMD64.exe
> .\mkdwarfs.exe --help

Alternatively, you can select the tool by passing --tool=<name> as
the first argument on the command line:

> .\dwarfs-universal-0.7.0-Windows-AMD64.exe --tool=mkdwarfs --help

Note that just like the dwarfs.exe Windows binary, the universal
Windows binary depends on the winfsp-x64.dll from the
WinFsp project. However, for the
universal binary, the DLL is loaded lazily, so you can still use all
other tools without the DLL.
See the Windows Support section for more details.

Dependencies

DwarFS uses CMake as a build tool.

It uses both Boost and
Folly, though the latter is
included as a submodule since very few distributions actually
offer packages for it. Folly itself has a number of dependencies,
so please check here
for an up-to-date list.

It also uses Facebook Thrift,
in particular the frozen library, for storing metadata in a highly
space-efficient, memory-mappable and well defined format. It’s also
included as a submodule, and we only build the compiler and a very
reduced library that contains just enough for DwarFS to work.

Other than that, DwarFS really only depends on FUSE3 and on a set
of compression libraries that Folly already depends on (namely
lz4, zstd
and liblzma).

The dependency on googletest
will be automatically resolved if you build with tests.

A good starting point for apt-based systems is probably:

$ apt install \
    gcc \
    g++ \
    clang \
    git \
    ccache \
    ninja-build \
    cmake \
    make \
    bison \
    flex \
    ronn \
    fuse3 \
    pkg-config \
    binutils-dev \
    libacl1-dev \
    libarchive-dev \
    libbenchmark-dev \
    libboost-chrono-dev \
    libboost-context-dev \
    libboost-filesystem-dev \
    libboost-iostreams-dev \
    libboost-program-options-dev \
    libboost-regex-dev \
    libboost-system-dev \
    libboost-thread-dev \
    libbrotli-dev \
    libevent-dev \
    libhowardhinnant-date-dev \
    libjemalloc-dev \
    libdouble-conversion-dev \
    libiberty-dev \
    liblz4-dev \
    liblzma-dev \
    libmagic-dev \
    librange-v3-dev \
    libssl-dev \
    libunwind-dev \
    libdwarf-dev \
    libelf-dev \
    libfmt-dev \
    libfuse3-dev \
    libgoogle-glog-dev \
    libutfcpp-dev \
    libflac++-dev \
    python3-mistletoe

Note that when building with gcc, the optimization level will be
set to -O2 instead of the CMake default of -O3 for release
builds. At least with versions up to gcc-10, the -O3 build is
up to 70% slower than a
build with -O2.

Building

Firstly, either clone the repository…

$ git clone --recurse-submodules https://github.com/mhx/dwarfs
$ cd dwarfs

…or unpack the release archive:

$ tar xvf dwarfs-x.y.z.tar.bz2
$ cd dwarfs-x.y.z

Once all dependencies have been installed, you can build DwarFS
using:

$ mkdir build
$ cd build
$ cmake .. -DWITH_TESTS=1
$ make -j$(nproc)

You can then run tests with:

$ make test

All binaries use jemalloc
as a memory allocator by default, as it is typically uses much less
system memory compared to the glibc or tcmalloc allocators.
To disable the use of jemalloc, pass -DUSE_JEMALLOC=0 on the
cmake command line.

Installing

Installing is as easy as:

$ sudo make install

Though you don’t have to install the tools to play with them.

Static Builds

Attempting to build statically linked binaries is highly discouraged
and not officially supported. That being said, here’s how to set up
an environment where you might be able to build static binaries.

This has been tested with ubuntu-22.04-live-server-amd64.iso. First,
install all the packages listed as dependencies above. Also install:

$ apt install ccache ninja libacl1-dev

ccache and ninja are optional, but help with a speedy compile.

Depending on your distibution, you’ll need to build and install static
versions of some libraries, e.g. libarchive and libmagic for Ubuntu:

$ wget https://github.com/libarchive/libarchive/releases/download/v3.6.2/libarchive-3.6.2.tar.xz
$ tar xf libarchive-3.6.2.tar.xz && cd libarchive-3.6.2
$ ./configure --prefix=/opt/static-libs --without-iconv --without-xml2 --without-expat
$ make && sudo make install

$ wget ftp://ftp.astron.com/pub/file/file-5.44.tar.gz
$ tar xf file-5.44.tar.gz && cd file-5.44
$ ./configure --prefix=/opt/static-libs --enable-static=yes --enable-shared=no
$ make && make install

That’s it! Now you can try building static binaries for DwarFS:

$ git clone --recurse-submodules https://github.com/mhx/dwarfs
$ cd dwarfs && mkdir build && cd build
$ cmake .. -GNinja -DWITH_TESTS=1 -DSTATIC_BUILD_DO_NOT_USE=1 \
           -DSTATIC_BUILD_EXTRA_PREFIX=/opt/static-libs
$ ninja
$ ninja test

Usage

Please check out the manual pages for mkdwarfs,
dwarfs, dwarfsck and
dwarfsextract. You can also access the manual
pages using the --man option to each binary, e.g.:

$ mkdwarfs --man

The dwarfs manual page also shows an example for setting
up DwarFS with overlayfs
in order to create a writable file system mount on top a read-only
DwarFS image.

A description of the DwarFS filesystem format can be found in
dwarfs-format.

A high-level overview of the internal operation of mkdwarfs is shown
in this sequence diagram.

Windows Support

Support for the Windows operating system is currently experimental.
Having worked pretty much exclusively in a Unix world for the past two
decades, my experience with Windows development is rather limited and
I’d expect there to definitely be bugs and rough edges in the Windows
code.

The Windows version of the DwarFS filesystem driver relies on the awesome
WinFsp project and its winfsp-x64.dll
must be discoverable by the dwarfs.exe driver.

The different tools should behave pretty much the same whether you’re
using them on Linux or Windows. The file system images can be copied
between Linux and Windows and images created on one OS should work fine
on the other.

There are a few things worth pointing out, though:

  • DwarFS supports both hardlinks and symlinks on Windows, just as it
    does on Linux. However, creating hardlinks and symlinks seems to
    require admin privileges on Windows, so if you want to e.g. extract
    a DwarFS image that contains links of some sort, you might run into
    errors if you don’t have the right privileges.

  • Due to a problem in
    WinFsp, symlinks cannot currently point outside of the mounted file
    system. Furthermore, due to another
    problem in WinFsp,
    symlinks with a drive letter will appear with a mangled target path.

  • The DwarFS driver on Windows correctly reports hardlink counts via
    its API, but currently these counts are not correctly propagated
    to the Windows file system layer. This is presumably due to a
    problem in WinFsp.

  • When mounting a DwarFS image on Windows, the mount point must not
    exist. This is different from Linux, where the mount point must
    actually exist. Also, it’s possible to mount a DwarFS image as a
    drive letter, e.g.

    dwarfs.exe image.dwarfs Z:

  • Filter rules for mkdwarfs always require Unix path separators,
    regardless of whether it’s running on Windows or Linux.

Building on Windows

Building on Windows is not too complicated thanks to vcpkg.
You’ll need to install:

WinFsp is expected to be installed in C:\Program Files (x68)\WinFsp;
if it’s not, you’ll need to set WINFSP_PATH when running CMake via
cmake/win.bat.

Now you need to clone vcpkg and dwarfs:

> cd %HOMEPATH%
> mkdir git
> cd git
> git clone https://github.com/Microsoft/vcpkg.git
> git clone https://github.com/mhx/dwarfs

Then, bootstrap vcpkg:

> .\vcpkg\bootstrap-vcpkg.bat

And build DwarFS:

> cd dwarfs
> mkdir build
> cd build
> ..\cmake\win.bat
> ninja

Once that’s done, you should be able to run the tests.
Set CTEST_PARALLEL_LEVEL according to the number of CPU cores in
your machine.

> set CTEST_PARALLEL_LEVEL=10
> ninja test

macOS Support

Support for the macOS operating system is currently experimental.

The macOS version of the DwarFS filesystem driver relies on the awesome
macFUSE project.

Building on macOS

Building on macOS involves a few steps, but should be relatively
straightforward:

  • Install Homebrew

  • Use Homebrew to install the necessary dependencies:

$ brew install cmake ninja ronn macfuse python3 brotli howard-hinnant-date \
               double-conversion fmt glog libarchive libevent flac openssl \
               pkg-config range-v3 utf8cpp xxhash boost zstd jemalloc

  • When installing macFUSE for the first time, you’ll need to explicitly
    allow the sofware in System Preferences / Privacy & Security. It’s
    quite likely that you’ll have to reboot after this.

  • Clone the DwarFS repository:

$ git clone --recurse-submodules https://github.com/mhx/dwarfs
  • Prepare the build by installing the mistletoe python module
    in a virtualenv:
$ cd dwarfs
$ python3 -m venv @buildenv
$ source ./@buildenv/bin/activate
$ pip3 install mistletoe
  • Build DwarFS and run its tests:
$ git checkout v0.9.4
$ git submodule update
$ mkdir build && cd build
$ cmake .. -GNinja -DWITH_TESTS=ON
$ ninja
$ export CTEST_PARALLEL_LEVEL=$(sysctl -n hw.logicalcpu)
$ ninja test
  • Install DwarFS:
$ ninja install

That’s it!

Use Cases

Astrophotography

Astrophotography can generate huge amounts of raw image data. During a
single night, it’s not unlikely to end up with a few dozens of gigabytes
of data. With most dedicated astrophotography cameras, this data ends up
in the form of FITS images. These are usually uncompressed, don’t compress
very well with standard compression algorithms, and while there are certain
compressed FITS formats, these aren’t widely supported.

One of the compression formats (simply called “Rice”) compresses reasonably
well and is really fast. However, its implementation for compressed FITS
has a few drawbacks. The most severe drawbacks are that compression isn’t
quite as good as it could be for color sensors and sensors with a less than
16 bits of resolution.

DwarFS supports the ricepp (Rice++) compression, which builds on the basic
idea of Rice compression, but makes a few enhancements: it compresses color
and low bit depth images significantly better and always searches for the
optimum solution during compression instead of relying on a heuristic.

Let’s look at an example using 129 images (darks, flats and lights) taken
with an ASI1600MM camera. Each image is 32 MiB, so a total of 4 GiB of data.
Compressing these with the standard fpack tool takes about 16.6 seconds
and yields a total output size of 2.2 GiB:

$ time fpack */*.fit */*/*.fit

user	14.992
system	1.592
total	16.616

$ find . -name '*.fz' -print0 | xargs -0 cat | wc -c
2369943360

However, this leaves you with *.fz files that not every application can
actually read.

Using DwarFS, here’s what we get:

$ mkdwarfs -i ASI1600 -o asi1600-20.dwarfs -S 20 --categorize
I 08:47:47.459077 scanning "ASI1600"
I 08:47:47.491492 assigning directory and link inodes...
I 08:47:47.491560 waiting for background scanners...
I 08:47:47.675241 scanning CPU time: 1.051s
I 08:47:47.675271 finalizing file inodes...
I 08:47:47.675330 saved 0 B / 3.941 GiB in 0/258 duplicate files
I 08:47:47.675360 assigning device inodes...
I 08:47:47.675371 assigning pipe/socket inodes...
I 08:47:47.675381 building metadata...
I 08:47:47.675393 building blocks...
I 08:47:47.675398 saving names and symlinks...
I 08:47:47.675514 updating name and link indices...
I 08:47:47.675796 waiting for segmenting/blockifying to finish...
I 08:47:50.274285 total ordering CPU time: 616.3us
I 08:47:50.274329 total segmenting CPU time: 1.132s
I 08:47:50.279476 saving chunks...
I 08:47:50.279622 saving directories...
I 08:47:50.279674 saving shared files table...
I 08:47:50.280745 saving names table... [1.047ms]
I 08:47:50.280768 saving symlinks table... [743ns]
I 08:47:50.282031 waiting for compression to finish...
I 08:47:50.823924 compressed 3.941 GiB to 1.201 GiB (ratio=0.304825)
I 08:47:50.824280 compression CPU time: 17.92s
I 08:47:50.824316 filesystem created without errors [3.366s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
waiting for block compression to finish
5 dirs, 0/0 soft/hard links, 258/258 files, 0 other
original size: 3.941 GiB, hashed: 315.4 KiB (18 files, 0 B/s)
scanned: 3.941 GiB (258 files, 117.1 GiB/s), categorizing: 0 B/s
saved by deduplication: 0 B (0 files), saved by segmenting: 0 B
filesystem: 3.941 GiB in 4037 blocks (4550 chunks, 516/516 fragments, 258 inodes)
compressed filesystem: 4037 blocks/1.201 GiB written

In less than 3.4 seconds, it compresses the data down to 1.2 GiB, almost
half the size of the fpack output.

In addition to saving a lot of disk space, this can also be useful when your
data is stored on a NAS. Here’s a comparison of the same set of data accessed
over a 1 Gb/s network connection, first using the uncompressed raw data:

find /mnt/ASI1600 -name '*.fit' -print0 | xargs -0 -P4 -n1 cat | dd of=/dev/null status=progress
4229012160 bytes (4.2 GB, 3.9 GiB) copied, 36.0455 s, 117 MB/s

And next using a DwarFS image on the same share:

$ dwarfs /mnt/asi1600-20.dwarfs mnt

$ find mnt -name '*.fit' -print0 | xargs -0 -P4 -n1 cat | dd of=/dev/null status=progress
4229012160 bytes (4.2 GB, 3.9 GiB) copied, 14.3681 s, 294 MB/s

That’s roughly 2.5 times faster. You can very likely see similar results
with slow external hard drives.

Dealing with Bit Rot

Currently, DwarFS has no built-in ability to add recovery information to a
file system image. However, for archival purposes, it’s a good idea to have
such recovery infomation in order to be able to repair a damaged image.

This is fortunately relatively straightforward using something like
par2cmdline:

$ par2create -n1 asi1600-20.dwarfs

This will create two additional files that you can place alongside the image
(or on a different storage), as you’ll only need them if DwarFS has detected
an issue with the file system image. If there’s an issue, you can run

$ par2repair asi1600-20.dwarfs

which will very likely be able to recover the image if less than 5% (that’s
the default used by par2create) of the image are damaged.

Extended Attributes

Preserving Extended Attributes in DwarFS Images

Extended attributes are not currently supported. Any extended attributes
stored in the source file system will not currently be preserved when
building a DwarFS image using mkdwarfs.

Extended Attributes exposed by the FUSE Driver

That being said, the root inode of a mounted DwarFS image currently exposes
one or two extended attributes on Linux:

$ attr -l mnt
Attribute "dwarfs.driver.pid" has a 4 byte value for mnt
Attribute "dwarfs.driver.perfmon" has a 4849 byte value for mnt

The dwarfs.driver.pid attribute simply contains the PID of the DwarFS
FUSE driver. The dwarfs.driver.perfmon attribute contains the current
results of the performance monitor.

Furthermore, each regular file exposes an attribute dwarfs.inodeinfo
with information about the undelying inode:

$ attr -l "05 Disappear.caf"
Attribute "dwarfs.inodeinfo" has a 448 byte value for 05 Disappear.caf

The attribute contains a JSON object with information about the
underlying inode:

$ attr -qg dwarfs.inodeinfo "05 Disappear.caf"
{
  "chunks": [
    {
      "block": 2,
      "category": "pcmaudio/metadata",
      "offset": 270976,
      "size": 4096
    },
    {
      "block": 414,
      "category": "pcmaudio/waveform",
      "offset": 37594368,
      "size": 29514492
    },
    {
      "block": 419,
      "category": "pcmaudio/waveform",
      "offset": 0,
      "size": 29385468
    }
  ],
  "gid": 100,
  "mode": 33188,
  "modestring": "----rw-r--r--",
  "uid": 1000
}

This is useful, for example, to check how a particular file is spread
across multiple blocks or which categories have been assigned to the
file.

Comparison

The SquashFS, xz, lrzip, zpaq and wimlib tests were all done on
an 8 core Intel® Xeon® E-2286M CPU @ 2.40GHz with 64 GiB of RAM.

The Cromfs tests were done with an older version of DwarFS
on a 6 core Intel® Xeon® CPU D-1528 @ 1.90GHz with 64 GiB of RAM.

The EROFS tests were done using DwarFS v0.9.8 and EROFS v1.7.1 on an
Intel® Core™ i9-13900K with 64 GiB of RAM.

The systems were mostly idle during all of the tests.

With SquashFS

The source directory contained 1139 different Perl installations
from 284 distinct releases, a total of 47.65 GiB of data in 1,927,501
files and 330,733 directories. The source directory was freshly
unpacked from a tar archive to an XFS partition on a 970 EVO Plus 2TB
NVME drive, so most of its contents were likely cached.

I’m using the same compression type and compression level for
SquashFS that is the default setting for DwarFS:

$ time mksquashfs install perl-install.squashfs -comp zstd -Xcompression-level 22
Parallel mksquashfs: Using 16 processors
Creating 4.0 filesystem on perl-install-zstd.squashfs, block size 131072.
[=========================================================/] 2107401/2107401 100%

Exportable Squashfs 4.0 filesystem, zstd compressed, data block size 131072
        compressed data, compressed metadata, compressed fragments,
        compressed xattrs, compressed ids
        duplicates are removed
Filesystem size 4637597.63 Kbytes (4528.90 Mbytes)
        9.29% of uncompressed filesystem size (49922299.04 Kbytes)
Inode table size 19100802 bytes (18653.13 Kbytes)
        26.06% of uncompressed inode table size (73307702 bytes)
Directory table size 19128340 bytes (18680.02 Kbytes)
        46.28% of uncompressed directory table size (41335540 bytes)
Number of duplicate files found 1780387
Number of inodes 2255794
Number of files 1925061
Number of fragments 28713
Number of symbolic links  0
Number of device nodes 0
Number of fifo nodes 0
Number of socket nodes 0
Number of directories 330733
Number of ids (unique uids + gids) 2
Number of uids 1
        mhx (1000)
Number of gids 1
        users (100)

real    32m54.713s
user    501m46.382s
sys     0m58.528s

For DwarFS, I’m sticking to the defaults:

$ time mkdwarfs -i install -o perl-install.dwarfs
I 11:33:33.310931 scanning install
I 11:33:39.026712 waiting for background scanners...
I 11:33:50.681305 assigning directory and link inodes...
I 11:33:50.888441 finding duplicate files...
I 11:34:01.120800 saved 28.2 GiB / 47.65 GiB in 1782826/1927501 duplicate files
I 11:34:01.122608 waiting for inode scanners...
I 11:34:12.839065 assigning device inodes...
I 11:34:12.875520 assigning pipe/socket inodes...
I 11:34:12.910431 building metadata...
I 11:34:12.910524 building blocks...
I 11:34:12.910594 saving names and links...
I 11:34:12.910691 bloom filter size: 32 KiB
I 11:34:12.910760 ordering 144675 inodes using nilsimsa similarity...
I 11:34:12.915555 nilsimsa: depth=20000 (1000), limit=255
I 11:34:13.052525 updating name and link indices...
I 11:34:13.276233 pre-sorted index (660176 name, 366179 path lookups) [360.6ms]
I 11:35:44.039375 144675 inodes ordered [91.13s]
I 11:35:44.041427 waiting for segmenting/blockifying to finish...
I 11:37:38.823902 bloom filter reject rate: 96.017% (TPR=0.244%, lookups=4740563665)
I 11:37:38.823963 segmentation matches: good=454708, bad=6819, total=464247
I 11:37:38.824005 segmentation collisions: L1=0.008%, L2=0.000% [2233254 hashes]
I 11:37:38.824038 saving chunks...
I 11:37:38.860939 saving directories...
I 11:37:41.318747 waiting for compression to finish...
I 11:38:56.046809 compressed 47.65 GiB to 430.9 MiB (ratio=0.00883101)
I 11:38:56.304922 filesystem created without errors [323s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
waiting for block compression to finish
330733 dirs, 0/2440 soft/hard links, 1927501/1927501 files, 0 other
original size: 47.65 GiB, dedupe: 28.2 GiB (1782826 files), segment: 15.19 GiB
filesystem: 4.261 GiB in 273 blocks (319178 chunks, 144675/144675 inodes)
compressed filesystem: 273 blocks/430.9 MiB written [depth: 20000]
█████████████████████████████████████████████████████████████████████████████▏100% |

real    5m23.030s
user    78m7.554s
sys     1m47.968s

So in this comparison, mkdwarfs is more than 6 times faster than mksquashfs,
both in terms of CPU time and wall clock time.

$ ll perl-install.*fs
-rw-r--r-- 1 mhx users  447230618 Mar  3 20:28 perl-install.dwarfs
-rw-r--r-- 1 mhx users 4748902400 Mar  3 20:10 perl-install.squashfs

In terms of compression ratio, the DwarFS file system is more than 10 times
smaller than the SquashFS file system. With DwarFS, the content has been
compressed down to less than 0.9% (!) of its original size. This compression
ratio only considers the data stored in the individual files, not the actual
disk space used. On the original XFS file system, according to du, the
source folder uses 52 GiB, so the DwarFS image actually only uses 0.8% of
the original space.

Here’s another comparison using lzma compression instead of zstd:

$ time mksquashfs install perl-install-lzma.squashfs -comp lzma

real    13m42.825s
user    205m40.851s
sys     3m29.088s

$ time mkdwarfs -i install -o perl-install-lzma.dwarfs -l9

real    3m43.937s
user    49m45.295s
sys     1m44.550s

$ ll perl-install-lzma.*fs
-rw-r--r-- 1 mhx users  315482627 Mar  3 21:23 perl-install-lzma.dwarfs
-rw-r--r-- 1 mhx users 3838406656 Mar  3 20:50 perl-install-lzma.squashfs

It’s immediately obvious that the runs are significantly faster and the
resulting images are significantly smaller. Still, mkdwarfs is about
4 times faster and produces and image that’s 12 times smaller than
the SquashFS image. The DwarFS image is only 0.6% of the original file size.

So, why not use lzma instead of zstd by default? The reason is that lzma
is about an order of magnitude slower to decompress than zstd. If you’re
only accessing data on your compressed filesystem occasionally, this might
not be a big deal, but if you use it extensively, zstd will result in
better performance.

The comparisons above are not completely fair. mksquashfs by default
uses a block size of 128KiB, whereas mkdwarfs uses 16MiB blocks by default,
or even 64MiB blocks with -l9. When using identical block sizes for both
file systems, the difference, quite expectedly, becomes a lot less dramatic:

$ time mksquashfs install perl-install-lzma-1M.squashfs -comp lzma -b 1M

real    15m43.319s
user    139m24.533s
sys     0m45.132s

$ time mkdwarfs -i install -o perl-install-lzma-1M.dwarfs -l9 -S20 -B3

real    4m25.973s
user    52m15.100s
sys     7m41.889s

$ ll perl-install*.*fs
-rw-r--r-- 1 mhx users  935953866 Mar 13 12:12 perl-install-lzma-1M.dwarfs
-rw-r--r-- 1 mhx users 3407474688 Mar  3 21:54 perl-install-lzma-1M.squashfs

Even this is still not entirely fair, as it uses a feature (-B3) that allows
DwarFS to reference file chunks from up to two previous filesystem blocks.

But the point is that this is really where SquashFS tops out, as it doesn’t
support larger block sizes or back-referencing. And as you’ll see below, the
larger blocks that DwarFS is using by default don’t necessarily negatively
impact performance.

DwarFS also features an option to recompress an existing file system with
a different compression algorithm. This can be useful as it allows relatively
fast experimentation with different algorithms and options without requiring
a full rebuild of the file system. For example, recompressing the above file
system with the best possible compression (-l 9):

$ time mkdwarfs --recompress -i perl-install.dwarfs -o perl-lzma-re.dwarfs -l9
I 20:28:03.246534 filesystem rewrittenwithout errors [148.3s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
filesystem: 4.261 GiB in 273 blocks (0 chunks, 0 inodes)
compressed filesystem: 273/273 blocks/372.7 MiB written
████████████████████████████████████████████████████████████████████▏100% \

real    2m28.279s
user    37m8.825s
sys     0m43.256s

$ ll perl-*.dwarfs
-rw-r--r-- 1 mhx users 447230618 Mar  3 20:28 perl-install.dwarfs
-rw-r--r-- 1 mhx users 390845518 Mar  4 20:28 perl-lzma-re.dwarfs
-rw-r--r-- 1 mhx users 315482627 Mar  3 21:23 perl-install-lzma.dwarfs

Note that while the recompressed filesystem is smaller than the original image,
it is still a lot bigger than the filesystem we previously build with -l9.
The reason is that the recompressed image still uses the same block size, and
the block size cannot be changed by recompressing.

In terms of how fast the file system is when using it, a quick test
I’ve done is to freshly mount the filesystem created above and run
each of the 1139 perl executables to print their version.

$ hyperfine -c "umount mnt" -p "umount mnt; dwarfs perl-install.dwarfs mnt -o cachesize=1g -o workers=4; sleep 1" -P procs 5 20 -D 5 "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P{procs} sh -c '\$0 -v >/dev/null'"
Benchmark #1: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P5 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      1.810 s ±  0.013 s    [User: 1.847 s, System: 0.623 s]
  Range (min … max):    1.788 s …  1.825 s    10 runs

Benchmark #2: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      1.333 s ±  0.009 s    [User: 1.993 s, System: 0.656 s]
  Range (min … max):    1.321 s …  1.354 s    10 runs

Benchmark #3: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P15 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      1.181 s ±  0.018 s    [User: 2.086 s, System: 0.712 s]
  Range (min … max):    1.165 s …  1.214 s    10 runs

Benchmark #4: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      1.149 s ±  0.015 s    [User: 2.128 s, System: 0.781 s]
  Range (min … max):    1.136 s …  1.186 s    10 runs

These timings are for initial runs on a freshly mounted file system,
running 5, 10, 15 and 20 processes in parallel. 1.1 seconds means that
it takes only about 1 millisecond per Perl binary.

Following are timings for subsequent runs, both on DwarFS (at mnt)
and the original XFS (at install). DwarFS is around 15% slower here:

$ hyperfine -P procs 10 20 -D 10 -w1 "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P{procs} sh -c '\$0 -v >/dev/null'" "ls -1 install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P{procs} sh -c '\$0 -v >/dev/null'"
Benchmark #1: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):     347.0 ms ±   7.2 ms    [User: 1.755 s, System: 0.452 s]
  Range (min … max):   341.3 ms … 365.2 ms    10 runs

Benchmark #2: ls -1 install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):     302.5 ms ±   3.3 ms    [User: 1.656 s, System: 0.377 s]
  Range (min … max):   297.1 ms … 308.7 ms    10 runs

Benchmark #3: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):     342.2 ms ±   4.1 ms    [User: 1.766 s, System: 0.451 s]
  Range (min … max):   336.0 ms … 349.7 ms    10 runs

Benchmark #4: ls -1 install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):     302.0 ms ±   3.0 ms    [User: 1.659 s, System: 0.374 s]
  Range (min … max):   297.0 ms … 305.4 ms    10 runs

Summary
  'ls -1 install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null'' ran
    1.00 ± 0.01 times faster than 'ls -1 install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null''
    1.13 ± 0.02 times faster than 'ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null''
    1.15 ± 0.03 times faster than 'ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null''

Using the lzma-compressed file system, the metrics for initial runs look
considerably worse (about an order of magnitude):

$ hyperfine -c "umount mnt" -p "umount mnt; dwarfs perl-install-lzma.dwarfs mnt -o cachesize=1g -o workers=4; sleep 1" -P procs 5 20 -D 5 "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P{procs} sh -c '\$0 -v >/dev/null'"
Benchmark #1: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P5 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):     10.660 s ±  0.057 s    [User: 1.952 s, System: 0.729 s]
  Range (min … max):   10.615 s … 10.811 s    10 runs

Benchmark #2: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P10 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      9.092 s ±  0.021 s    [User: 1.979 s, System: 0.680 s]
  Range (min … max):    9.059 s …  9.126 s    10 runs

Benchmark #3: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P15 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      9.012 s ±  0.188 s    [User: 2.077 s, System: 0.702 s]
  Range (min … max):    8.839 s …  9.277 s    10 runs

Benchmark #4: ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '$0 -v >/dev/null'
  Time (mean ± σ):      9.004 s ±  0.298 s    [User: 2.134 s, System: 0.736 s]
  Range (min … max):    8.611 s …  9.555 s    10 runs

So you might want to consider using zstd instead of lzma if you’d
like to optimize for file system performance. It’s also the default
compression used by mkdwarfs.

Now here’s a comparison with the SquashFS filesystem:

$ hyperfine -c 'sudo umount mnt' -p 'umount mnt; dwarfs perl-install.dwarfs mnt -o cachesize=1g -o workers=4; sleep 1' -n dwarfs-zstd "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '\$0 -v >/dev/null'" -p 'sudo umount mnt; sudo mount -t squashfs perl-install.squashfs mnt; sleep 1' -n squashfs-zstd "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '\$0 -v >/dev/null'"
Benchmark #1: dwarfs-zstd
  Time (mean ± σ):      1.151 s ±  0.015 s    [User: 2.147 s, System: 0.769 s]
  Range (min … max):    1.118 s …  1.174 s    10 runs

Benchmark #2: squashfs-zstd
  Time (mean ± σ):      6.733 s ±  0.007 s    [User: 3.188 s, System: 17.015 s]
  Range (min … max):    6.721 s …  6.743 s    10 runs

Summary
  'dwarfs-zstd' ran
    5.85 ± 0.08 times faster than 'squashfs-zstd'

So, DwarFS is almost six times faster than SquashFS. But what’s more,
SquashFS also uses significantly more CPU power. However, the numbers
shown above for DwarFS obviously don’t include the time spent in the
dwarfs process, so I repeated the test outside of hyperfine:

$ time dwarfs perl-install.dwarfs mnt -o cachesize=1g -o workers=4 -f

real    0m4.569s
user    0m2.154s
sys     0m1.846s

So, in total, DwarFS was using 5.7 seconds of CPU time, whereas
SquashFS was using 20.2 seconds, almost four times as much. Ignore
the ‘real’ time, this is only how long it took me to unmount the
file system again after mounting it.

Another real-life test was to build and test a Perl module with 624
different Perl versions in the compressed file system. The module I’ve
used, Tie::Hash::Indexed,
has an XS component that requires a C compiler to build. So this really
accesses a lot of different stuff in the file system:

  • The perl executables and its shared libraries

  • The Perl modules used for writing the Makefile

  • Perl’s C header files used for building the module

  • More Perl modules used for running the tests

I wrote a little script to be able to run multiple builds in parallel:

#!/bin/bash
set -eu
perl=$1
dir=$(echo "$perl" | cut -d/ --output-delimiter=- -f5,6)
rsync -a Tie-Hash-Indexed/ $dir/
cd $dir
$1 Makefile.PL >/dev/null 2>&1
make test >/dev/null 2>&1
cd ..
rm -rf $dir
echo $perl

The following command will run up to 16 builds in parallel on the 8 core
Xeon CPU, including debug, optimized and threaded versions of all Perl
releases between 5.10.0 and 5.33.3, a total of 624 perl installations:

$ time ls -1 /tmp/perl/install/*/perl-5.??.?/bin/perl5* | sort -t / -k 8 | xargs -d $'\n' -P 16 -n 1 ./build.sh

Tests were done with a cleanly mounted file system to make sure the caches
were empty. ccache was primed to make sure all compiler runs could be
satisfied from the cache. With SquashFS, the timing was:

real    0m52.385s
user    8m10.333s
sys     4m10.056s

And with DwarFS:

real    0m50.469s
user    9m22.597s
sys     1m18.469s

So, frankly, not much of a difference, with DwarFS being just a bit faster.
The dwarfs process itself used:

real    0m56.686s
user    0m18.857s
sys     0m21.058s

So again, DwarFS used less raw CPU power overall, but in terms of wallclock
time, the difference is really marginal.

With SquashFS & xz

This test uses slightly less pathological input data: the root filesystem of
a recent Raspberry Pi OS release. This file system also contains device inodes,
so in order to preserve those, we pass --with-devices to mkdwarfs:

$ time sudo mkdwarfs -i raspbian -o raspbian.dwarfs --with-devices
I 21:30:29.812562 scanning raspbian
I 21:30:29.908984 waiting for background scanners...
I 21:30:30.217446 assigning directory and link inodes...
I 21:30:30.221941 finding duplicate files...
I 21:30:30.288099 saved 31.05 MiB / 1007 MiB in 1617/34582 duplicate files
I 21:30:30.288143 waiting for inode scanners...
I 21:30:31.393710 assigning device inodes...
I 21:30:31.394481 assigning pipe/socket inodes...
I 21:30:31.395196 building metadata...
I 21:30:31.395230 building blocks...
I 21:30:31.395291 saving names and links...
I 21:30:31.395374 ordering 32965 inodes using nilsimsa similarity...
I 21:30:31.396254 nilsimsa: depth=20000 (1000), limit=255
I 21:30:31.407967 pre-sorted index (46431 name, 2206 path lookups) [11.66ms]
I 21:30:31.410089 updating name and link indices...
I 21:30:38.178505 32965 inodes ordered [6.783s]
I 21:30:38.179417 waiting for segmenting/blockifying to finish...
I 21:31:06.248304 saving chunks...
I 21:31:06.251998 saving directories...
I 21:31:06.402559 waiting for compression to finish...
I 21:31:16.425563 compressed 1007 MiB to 287 MiB (ratio=0.285036)
I 21:31:16.464772 filesystem created without errors [46.65s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
waiting for block compression to finish
4435 dirs, 5908/0 soft/hard links, 34582/34582 files, 7 other
original size: 1007 MiB, dedupe: 31.05 MiB (1617 files), segment: 47.23 MiB
filesystem: 928.4 MiB in 59 blocks (38890 chunks, 32965/32965 inodes)
compressed filesystem: 59 blocks/287 MiB written [depth: 20000]
████████████████████████████████████████████████████████████████████▏100% |

real    0m46.711s
user    10m39.038s
sys     0m8.123s

Again, SquashFS uses the same compression options:

$ time sudo mksquashfs raspbian raspbian.squashfs -comp zstd -Xcompression-level 22
Parallel mksquashfs: Using 16 processors
Creating 4.0 filesystem on raspbian.squashfs, block size 131072.
[===============================================================\] 39232/39232 100%

Exportable Squashfs 4.0 filesystem, zstd compressed, data block size 131072
        compressed data, compressed metadata, compressed fragments,
        compressed xattrs, compressed ids
        duplicates are removed
Filesystem size 371934.50 Kbytes (363.22 Mbytes)
        35.98% of uncompressed filesystem size (1033650.60 Kbytes)
Inode table size 399913 bytes (390.54 Kbytes)
        26.53% of uncompressed inode table size (1507581 bytes)
Directory table size 408749 bytes (399.17 Kbytes)
        42.31% of uncompressed directory table size (966174 bytes)
Number of duplicate files found 1618
Number of inodes 44932
Number of files 34582
Number of fragments 3290
Number of symbolic links  5908
Number of device nodes 7
Number of fifo nodes 0
Number of socket nodes 0
Number of directories 4435
Number of ids (unique uids + gids) 18
Number of uids 5
        root (0)
        mhx (1000)
        unknown (103)
        shutdown (6)
        unknown (106)
Number of gids 15
        root (0)
        unknown (109)
        unknown (42)
        unknown (1000)
        users (100)
        unknown (43)
        tty (5)
        unknown (108)
        unknown (111)
        unknown (110)
        unknown (50)
        mail (12)
        nobody (65534)
        adm (4)
        mem (8)

real    0m50.124s
user    9m41.708s
sys     0m1.727s

The difference in speed is almost negligible. SquashFS is just a bit
slower here. In terms of compression, the difference also isn’t huge:

$ ls -lh raspbian.* *.xz
-rw-r--r-- 1 mhx  users 297M Mar  4 21:32 2020-08-20-raspios-buster-armhf-lite.img.xz
-rw-r--r-- 1 root root  287M Mar  4 21:31 raspbian.dwarfs
-rw-r--r-- 1 root root  364M Mar  4 21:33 raspbian.squashfs

Interestingly, xz actually can’t compress the whole original image
better than DwarFS.

We can even again try to increase the DwarFS compression level:

$ time sudo mkdwarfs -i raspbian -o raspbian-9.dwarfs --with-devices -l9

real    0m54.161s
user    8m40.109s
sys     0m7.101s

Now that actually gets the DwarFS image size well below that of the
xz archive:

$ ls -lh raspbian-9.dwarfs *.xz
-rw-r--r-- 1 root root  244M Mar  4 21:36 raspbian-9.dwarfs
-rw-r--r-- 1 mhx  users 297M Mar  4 21:32 2020-08-20-raspios-buster-armhf-lite.img.xz

Even if you actually build a tarball and compress that (instead of
compressing the EXT4 file system itself), xz isn’t quite able to
match the DwarFS image size:

$ time sudo tar cf - raspbian | xz -9 -vT 0 >raspbian.tar.xz
  100 %     246.9 MiB / 1,037.2 MiB = 0.238    13 MiB/s       1:18

real    1m18.226s
user    6m35.381s
sys     0m2.205s

$ ls -lh raspbian.tar.xz
-rw-r--r-- 1 mhx users 247M Mar  4 21:40 raspbian.tar.xz

DwarFS also comes with the dwarfsextract tool
that allows extraction of a filesystem image without the FUSE driver.
So here’s a comparison of the extraction speed:

$ time sudo tar xf raspbian.tar.xz -C out1

real    0m12.846s
user    0m12.313s
sys     0m1.616s

$ time sudo dwarfsextract -i raspbian-9.dwarfs -o out2

real    0m3.825s
user    0m13.234s
sys     0m1.382s

So, dwarfsextract is almost 4 times faster thanks to using multiple
worker threads for decompression. It’s writing about 300 MiB/s in this
example.

Another nice feature of dwarfsextract is that it allows you to directly
output data in an archive format, so you could create a tarball from
your image without extracting the files to disk:

$ dwarfsextract -i raspbian-9.dwarfs -f ustar | xz -9 -T0 >raspbian2.tar.xz

This has the interesting side-effect that the resulting tarball will
likely be smaller than the one built straight from the directory:

$ ls -lh raspbian*.tar.xz
-rw-r--r-- 1 mhx users 247M Mar  4 21:40 raspbian.tar.xz
-rw-r--r-- 1 mhx users 240M Mar  4 23:52 raspbian2.tar.xz

That’s because dwarfsextract writes files in inode-order, and by
default inodes are ordered by similarity for the best possible
compression.

With lrzip

lrzip is a compression utility
targeted especially at compressing large files. From its description,
it looks like it does something very similar to DwarFS, i.e. it looks
for duplicate segments before passsing the de-duplicated data on to
an lzma compressor.

When I first read about lrzip, I was pretty certain it would easily
beat DwarFS. So let’s take a look. lrzip operates on a single file,
so it’s necessary to first build a tarball:

$ time tar cf perl-install.tar install

real    2m9.568s
user    0m3.757s
sys     0m26.623s

Now we can run lrzip:

$ time lrzip -vL9 -o perl-install.tar.lrzip perl-install.tar
The following options are in effect for this COMPRESSION.
Threading is ENABLED. Number of CPUs detected: 16
Detected 67106172928 bytes ram
Compression level 9
Nice Value: 19
Show Progress
Verbose
Output Filename Specified: perl-install.tar.lrzip
Temporary Directory set as: ./
Compression mode is: LZMA. LZO Compressibility testing enabled
Heuristically Computed Compression Window: 426 = 42600MB
File size: 52615639040
Will take 2 passes
Beginning rzip pre-processing phase
Beginning rzip pre-processing phase
perl-install.tar - Compression Ratio: 100.378. Average Compression Speed: 14.536MB/s.
Total time: 00:57:32.47

real    57m32.472s
user    81m44.104s
sys     4m50.221s

That definitely took a while. This is about an order of magnitude
slower than mkdwarfs and it barely makes use of the 8 cores.

$ ll -h perl-install.tar.lrzip
-rw-r--r-- 1 mhx users 500M Mar  6 21:16 perl-install.tar.lrzip

This is a surprisingly disappointing result. The archive is 65% larger
than a DwarFS image at -l9 that takes less than 4 minutes to build.
Also, you can’t just access the files in the .lrzip without fully
unpacking the archive first.

That being said, it is better than just using xz on the tarball:

$ time xz -T0 -v9 -c perl-install.tar >perl-install.tar.xz
perl-install.tar (1/1)
  100 %      4,317.0 MiB / 49.0 GiB = 0.086    24 MiB/s      34:55

real    34m55.450s
user    543m50.810s
sys     0m26.533s

$ ll perl-install.tar.xz -h
-rw-r--r-- 1 mhx users 4.3G Mar  6 22:59 perl-install.tar.xz

With zpaq

zpaq is a journaling backup
utility and archiver. Again, it appears to share some of the ideas in
DwarFS, like segmentation analysis, but it also adds some features on
top that make it useful for incremental backups. However, it’s also
not usable as a file system, so data needs to be extracted before it
can be used.

Anyway, how does it fare in terms of speed and compression performance?

$ time zpaq a perl-install.zpaq install -m5

After a few million lines of output that (I think) cannot be turned off:

2258234 +added, 0 -removed.

0.000000 + (51161.953159 -> 8932.000297 -> 490.227707) = 490.227707 MB
2828.082 seconds (all OK)

real    47m8.104s
user    714m44.286s
sys     3m6.751s

So, it’s an order of magnitude slower than mkdwarfs and uses 14 times
as much CPU resources as mkdwarfs -l9. The resulting archive it pretty
close in size to the default configuration DwarFS image, but it’s more
than 50% bigger than the image produced by mkdwarfs -l9.

$ ll perl-install*.*
-rw-r--r-- 1 mhx users 490227707 Mar  7 01:38 perl-install.zpaq
-rw-r--r-- 1 mhx users 315482627 Mar  3 21:23 perl-install-l9.dwarfs
-rw-r--r-- 1 mhx users 447230618 Mar  3 20:28 perl-install.dwarfs

What’s really surprising is how slow it is to extract the zpaq
archive again:

$ time zpaq x perl-install.zpaq
2798.097 seconds (all OK)

real    46m38.117s
user    711m18.734s
sys     3m47.876s

That’s 700 times slower than extracting the DwarFS image.

With zpaqfranz

zpaqfranz is a derivative of zpaq.
Much to my delight, it doesn’t generate millions of lines of output.
It claims to be multi-threaded and de-duplicating, so definitely worth
taking a look. Like zpaq, it supports incremental backups.

We’ll use a different input to compare zpaqfranz and DwarFS: The source code
of 670 different releases of the “wine” emulator. That’s 73 gigabytes of data
in total, spread across slightly more than 3 million files. It’s obviously
highly redundant and should thus be a good data set to compare the tools.
For reference, a .tar.xz of the directory is still 7 GiB in size and a
SquashFS image of the data gets down to around 1.6 GiB. An “optimized”
.tar.xz, where the input files were ordered by similarity, compresses down
to 399 MiB, almost 20 times better than without ordering.

Now it’s time to try zpaqfranz. The input data is stored on a fast SSD and a
large fraction of it is already in the file system cache from previous runs,
so disk I/O is not a bottleneck.

$ time ./zpaqfranz a winesrc.zpaq winesrc
zpaqfranz v58.8k-JIT-L(2023-08-05)
Creating winesrc.zpaq at offset 0 + 0
Add 2024-01-11 07:25:22 3.117.413     69.632.090.852 (  64.85 GB) 16T (362.904 dirs)
3.480.317 +added, 0 -removed.

0 + (69.632.090.852 -> 8.347.553.798 -> 617.600.892) = 617.600.892 @ 58.38 MB/s

1137.441 seconds (000:18:57) (all OK)

real    18m58.632s
user    11m51.052s
sys     1m3.389s

That is considerably faster than the original zpaq, and uses about 60 times
less CPU resources. The output file is 589 MiB, so slightly larger than both
the “optimized” .tar.gz and the zpaq output.

How does mkdwarfs do?

$ time mkdwarfs -i winesrc -o winesrc.dwarfs -l9
[...]
I 07:55:20.546636 compressed 64.85 GiB to 93.2 MiB (ratio=0.00140344)
I 07:55:20.826699 compression CPU time: 6.726m
I 07:55:20.827338 filesystem created without errors [2.283m]
[...]

real    2m17.100s
user    9m53.633s
sys     2m29.236s

It uses pretty much the same amount of CPU resources, but finishes more than
8 times faster. The DwarFS output file is more than 6 times smaller.

You can actually squeeze a bit more redundancy out of the original data by
tweaking the similarity ordering and switching from lzma to brotli compression,
albeit at a somewhat slower compression speed:

mkdwarfs -i winesrc -o winesrc.dwarfs -l9 -C brotli:quality=11:lgwin=26 --order=nilsimsa:max-cluster-size=200k
[...]
I 08:21:01.138075 compressed 64.85 GiB to 73.52 MiB (ratio=0.00110716)
I 08:21:01.485737 compression CPU time: 36.58m
I 08:21:01.486313 filesystem created without errors [5.501m]
[...]
real    5m30.178s
user    40m59.193s
sys     2m36.234s

That’s almost a 1000x reduction in size.

Let’s also look at decompression speed:

$ time zpaqfranz x winesrc.zpaq
zpaqfranz v58.8k-JIT-L(2023-08-05)
/home/mhx/winesrc.zpaq:
1 versions, 3.480.317 files, 617.600.892 bytes (588.99 MB)
Extract 69.632.090.852 bytes (64.85 GB) in 3.117.413 files (362.904 folders) / 16 T
        99.18% 00:00:00  (  64.32 GB)=>(  64.85 GB)  548.83 MB/sec

125.636 seconds (000:02:05) (all OK)

real    2m6.968s
user    1m36.177s
sys     1m10.980s

$ time dwarfsextract -i winesrc.dwarfs

real    1m49.182s
user    0m34.667s
sys     1m28.733s

Decompression time is pretty much in the same ballpark, with just slightly
shorter times for the DwarFS image.

With wimlib

wimlib is a really interesting project that is
a lot more mature than DwarFS. While DwarFS at its core has a library
component that could potentially be ported to other operating systems,
wimlib already is available on many platforms. It also seems to have
quite a rich set of features, so it’s definitely worth taking a look at.

I first tried wimcapture on the perl dataset:

$ time wimcapture --unix-data --solid --solid-chunk-size=16M install perl-install.wim
Scanning "install"
47 GiB scanned (1927501 files, 330733 directories)
Using LZMS compression with 16 threads
Archiving file data: 19 GiB of 19 GiB (100%) done

real    15m23.310s
user    174m29.274s
sys     0m42.921s

$ ll perl-install.*
-rw-r--r-- 1 mhx users  447230618 Mar  3 20:28 perl-install.dwarfs
-rw-r--r-- 1 mhx users  315482627 Mar  3 21:23 perl-install-l9.dwarfs
-rw-r--r-- 1 mhx users 4748902400 Mar  3 20:10 perl-install.squashfs
-rw-r--r-- 1 mhx users 1016981520 Mar  6 21:12 perl-install.wim

So, wimlib is definitely much better than squashfs, in terms of both
compression ratio and speed. DwarFS is however about 3 times faster to
create the file system and the DwarFS file system less than half the size.
When switching to LZMA compression, the DwarFS file system is more than
3 times smaller (wimlib uses LZMS compression by default).

What’s a bit surprising is that mounting a wim file takes quite a bit
of time:

$ time wimmount perl-install.wim mnt
[WARNING] Mounting a WIM file containing solid-compressed data; file access may be slow.

real    0m2.038s
user    0m1.764s
sys     0m0.242s

Mounting the DwarFS image takes almost no time in comparison:

$ time git/github/dwarfs/build-clang-11/dwarfs perl-install-default.dwarfs mnt
I 00:23:39.238182 dwarfs (v0.4.0, fuse version 35)

real    0m0.003s
user    0m0.003s
sys     0m0.000s

That’s just because it immediately forks into background by default and
initializes the file system in the background. However, even when
running it in the foreground, initializing the file system takes only
about 60 milliseconds:

$ dwarfs perl-install.dwarfs mnt -f
I 00:25:03.186005 dwarfs (v0.4.0, fuse version 35)
I 00:25:03.248061 file system initialized [60.95ms]

If you actually build the DwarFS file system with uncompressed metadata,
mounting is basically instantaneous:

$ dwarfs perl-install-meta.dwarfs mnt -f
I 00:27:52.667026 dwarfs (v0.4.0, fuse version 35)
I 00:27:52.671066 file system initialized [2.879ms]

I’ve tried running the benchmark where all 1139 perl executables
print their version with the wimlib image, but after about 10 minutes,
it still hadn’t finished the first run (with the DwarFS image, one run
took slightly more than 2 seconds). I then tried the following instead:

$ ls -1 /tmp/perl/install/*/*/bin/perl5* | xargs -d $'\n' -n1 -P1 sh -c 'time $0 -v >/dev/null' 2>&1 | grep ^real
real    0m0.802s
real    0m0.652s
real    0m1.677s
real    0m1.973s
real    0m1.435s
real    0m1.879s
real    0m2.003s
real    0m1.695s
real    0m2.343s
real    0m1.899s
real    0m1.809s
real    0m1.790s
real    0m2.115s

Judging from that, it would have probably taken about half an hour
for a single run, which makes at least the --solid wim image pretty
much unusable for actually working with the file system.

The --solid option was suggested to me because it resembles the way
that DwarFS actually organizes data internally. However, judging by the
warning when mounting a solid image, it’s probably not ideal when using
the image as a mounted file system. So I tried again without --solid:

$ time wimcapture --unix-data install perl-install-nonsolid.wim
Scanning "install"
47 GiB scanned (1927501 files, 330733 directories)
Using LZX compression with 16 threads
Archiving file data: 19 GiB of 19 GiB (100%) done

real    8m39.034s
user    64m58.575s
sys     0m32.003s

This is still more than 3 minutes slower than mkdwarfs. However, it
yields an image that’s almost 10 times the size of the DwarFS image
and comparable in size to the SquashFS image:

$ ll perl-install-nonsolid.wim -h
-rw-r--r-- 1 mhx users 4.6G Mar  6 23:24 perl-install-nonsolid.wim

This still takes surprisingly long to mount:

$ time wimmount perl-install-nonsolid.wim mnt

real    0m1.603s
user    0m1.327s
sys     0m0.275s

However, it’s really usable as a file system, even though it’s about
4-5 times slower than the DwarFS image:

$ hyperfine -c 'umount mnt' -p 'umount mnt; dwarfs perl-install.dwarfs mnt -o cachesize=1g -o workers=4; sleep 1' -n dwarfs "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '\$0 -v >/dev/null'" -p 'umount mnt; wimmount perl-install-nonsolid.wim mnt; sleep 1' -n wimlib "ls -1 mnt/*/*/bin/perl5* | xargs -d $'\n' -n1 -P20 sh -c '\$0 -v >/dev/null'"
Benchmark #1: dwarfs
  Time (mean ± σ):      1.149 s ±  0.019 s    [User: 2.147 s, System: 0.739 s]
  Range (min … max):    1.122 s …  1.187 s    10 runs

Benchmark #2: wimlib
  Time (mean ± σ):      7.542 s ±  0.069 s    [User: 2.787 s, System: 0.694 s]
  Range (min … max):    7.490 s …  7.732 s    10 runs

Summary
  'dwarfs' ran
    6.56 ± 0.12 times faster than 'wimlib'

With Cromfs

I used Cromfs in the past
for compressed file systems and remember that it did a pretty good job
in terms of compression ratio. But it was never fast. However, I didn’t
quite remember just how slow it was until I tried to set up a test.

Here’s a run on the Perl dataset, with the block size set to 16 MiB to
match the default of DwarFS, and with additional options suggested to
speed up compression:

$ time mkcromfs -f 16777216 -qq -e -r100000 install perl-install.cromfs
Writing perl-install.cromfs...
mkcromfs: Automatically enabling --24bitblocknums because it seems possible for this filesystem.
Root pseudo file is 108 bytes
Inotab spans 0x7f3a18259000..0x7f3a1bfffb9c
Root inode spans 0x7f3a205d2948..0x7f3a205d294c
Beginning task for Files and directories: Finding identical blocks
2163608 reuse opportunities found. 561362 unique blocks. Block table will be 79.4% smaller than without the index search.
Beginning task for Files and directories: Blockifying
Blockifying:  0.04% (140017/2724970) idx(siz=80423,del=0) rawin(20.97 MB)rawout(20.97 MB)diff(1956 bytes)
Termination signalled, cleaning up temporaries

real    29m9.634s
user    201m37.816s
sys     2m15.005s

So, it processed 21 MiB out of 48 GiB in half an hour, using almost
twice as much CPU resources as DwarFS for the whole file system.
At this point I decided it’s likely not worth waiting (presumably)
another month (!) for mkcromfs to finish. I double checked that
I didn’t accidentally build a debugging version, mkcromfs was
definitely built with -O3.

I then tried once more with a smaller version of the Perl dataset.
This only has 20 versions (instead of 1139) of Perl, and obviously
a lot less redundancy:

$ time mkcromfs -f 16777216 -qq -e -r100000 install-small perl-install.cromfs
Writing perl-install.cromfs...
mkcromfs: Automatically enabling --16bitblocknums because it seems possible for this filesystem.
Root pseudo file is 108 bytes
Inotab spans 0x7f00e0774000..0x7f00e08410a8
Root inode spans 0x7f00b40048f8..0x7f00b40048fc
Beginning task for Files and directories: Finding identical blocks
25362 reuse opportunities found. 9815 unique blocks. Block table will be 72.1% smaller than without the index search.
Beginning task for Files and directories: Blockifying
Compressing raw rootdir inode (28 bytes)z=982370,del=2) rawin(641.56 MB)rawout(252.72 MB)diff(388.84 MB)
 compressed into 35 bytes
INOTAB pseudo file is 839.85 kB
Inotab inode spans 0x7f00bc036ed8..0x7f00bc036ef4
Beginning task for INOTAB: Finding identical blocks
0 reuse opportunities found. 13 unique blocks. Block table will be 0.0% smaller than without the index search.
Beginning task for INOTAB: Blockifying
mkcromfs: Automatically enabling --packedblocks because it is possible for this filesystem.
Compressing raw inotab inode (52 bytes)
 compressed into 58 bytes
Compressing 9828 block records (4 bytes each, total 39312 bytes)
 compressed into 15890 bytes
Compressing and writing 16 fblocks...

16 fblocks were written: 35.31 MB = 13.90 % of 254.01 MB
Filesystem size: 35.33 MB = 5.50 % of original 642.22 MB
End

real    27m38.833s
user    277m36.208s
sys     11m36.945s

And repeating the same task with mkdwarfs:

$ time mkdwarfs -i install-small -o perl-install-small.dwarfs
21:13:38.131724 scanning install-small
21:13:38.320139 waiting for background scanners...
21:13:38.727024 assigning directory and link inodes...
21:13:38.731807 finding duplicate files...
21:13:38.832524 saved 267.8 MiB / 611.8 MiB in 22842/26401 duplicate files
21:13:38.832598 waiting for inode scanners...
21:13:39.619963 assigning device inodes...
21:13:39.620855 assigning pipe/socket inodes...
21:13:39.621356 building metadata...
21:13:39.621453 building blocks...
21:13:39.621472 saving names and links...
21:13:39.621655 ordering 3559 inodes using nilsimsa similarity...
21:13:39.622031 nilsimsa: depth=20000, limit=255
21:13:39.629206 updating name and link indices...
21:13:39.630142 pre-sorted index (3360 name, 2127 path lookups) [8.014ms]
21:13:39.752051 3559 inodes ordered [130.3ms]
21:13:39.752101 waiting for segmenting/blockifying to finish...
21:13:53.250951 saving chunks...
21:13:53.251581 saving directories...
21:13:53.303862 waiting for compression to finish...
21:14:11.073273 compressed 611.8 MiB to 24.01 MiB (ratio=0.0392411)
21:14:11.091099 filesystem created without errors [32.96s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
waiting for block compression to finish
3334 dirs, 0/0 soft/hard links, 26401/26401 files, 0 other
original size: 611.8 MiB, dedupe: 267.8 MiB (22842 files), segment: 121.5 MiB
filesystem: 222.5 MiB in 14 blocks (7177 chunks, 3559/3559 inodes)
compressed filesystem: 14 blocks/24.01 MiB written
██████████████████████████████████████████████████████████████████████▏100% \

real    0m33.007s
user    3m43.324s
sys     0m4.015s

So, mkdwarfs is about 50 times faster than mkcromfs and uses 75 times
less CPU resources. At the same time, the DwarFS file system is 30% smaller:

$ ls -l perl-install-small.*fs
-rw-r--r-- 1 mhx users 35328512 Dec  8 14:25 perl-install-small.cromfs
-rw-r--r-- 1 mhx users 25175016 Dec 10 21:14 perl-install-small.dwarfs

I noticed that the blockifying step that took ages for the full dataset
with mkcromfs ran substantially faster (in terms of MiB/second) on the
smaller dataset, which makes me wonder if there’s some quadratic complexity
behaviour that’s slowing down mkcromfs.

In order to be completely fair, I also ran mkdwarfs with -l 9 to enable
LZMA compression (which is what mkcromfs uses by default):

$ time mkdwarfs -i install-small -o perl-install-small-l9.dwarfs -l 9
21:16:21.874975 scanning install-small
21:16:22.092201 waiting for background scanners...
21:16:22.489470 assigning directory and link inodes...
21:16:22.495216 finding duplicate files...
21:16:22.611221 saved 267.8 MiB / 611.8 MiB in 22842/26401 duplicate files
21:16:22.611314 waiting for inode scanners...
21:16:23.394332 assigning device inodes...
21:16:23.395184 assigning pipe/socket inodes...
21:16:23.395616 building metadata...
21:16:23.395676 building blocks...
21:16:23.395685 saving names and links...
21:16:23.395830 ordering 3559 inodes using nilsimsa similarity...
21:16:23.396097 nilsimsa: depth=50000, limit=255
21:16:23.401042 updating name and link indices...
21:16:23.403127 pre-sorted index (3360 name, 2127 path lookups) [6.936ms]
21:16:23.524914 3559 inodes ordered [129ms]
21:16:23.525006 waiting for segmenting/blockifying to finish...
21:16:33.865023 saving chunks...
21:16:33.865883 saving directories...
21:16:33.900140 waiting for compression to finish...
21:17:10.505779 compressed 611.8 MiB to 17.44 MiB (ratio=0.0284969)
21:17:10.526171 filesystem created without errors [48.65s]
⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯⎯
waiting for block compression to finish
3334 dirs, 0/0 soft/hard links, 26401/26401 files, 0 other
original size: 611.8 MiB, dedupe: 267.8 MiB (22842 files), segment: 122.2 MiB
filesystem: 221.8 MiB in 4 blocks (7304 chunks, 3559/3559 inodes)
compressed filesystem: 4 blocks/17.44 MiB written
██████████████████████████████████████████████████████████████████████▏100% /

real    0m48.683s
user    2m24.905s
sys     0m3.292s

$ ls -l perl-install-small*.*fs
-rw-r--r-- 1 mhx users 18282075 Dec 10 21:17 perl-install-small-l9.dwarfs
-rw-r--r-- 1 mhx users 35328512 Dec  8 14:25 perl-install-small.cromfs
-rw-r--r-- 1 mhx users 25175016 Dec 10 21:14 perl-install-small.dwarfs

It takes about 15 seconds longer to build the DwarFS file system with LZMA
compression (this is still 35 times faster than Cromfs), but reduces the
size even further to make it almost half the size of the Cromfs file system.

I would have added some benchmarks with the Cromfs FUSE driver, but sadly
it crashed right upon trying to list the directory after mounting.

With EROFS

EROFS is a read-only compressed
file system that has been added to the Linux kernel recently.
Its goals are different from those of DwarFS, though. It is designed to
be lightweight (which DwarFS is definitely not) and to run on constrained
hardware like embedded devices or smartphones. It is not designed to provide
maximum compression. It currently supports LZ4 and LZMA compression.

Running it on the full Perl dataset using options given in the README for
“well-compressed images”:

$ time mkfs.erofs -C1048576 -Eztailpacking,fragments,all-fragments,dedupe -zlzma,9 perl-install-lzma9.erofs perl-install
mkfs.erofs 1.7.1-gd93a18c9
<W> erofs: It may take a longer time since MicroLZMA is still single-threaded for now.
Build completed.
------
Filesystem UUID: 538ce164-5f9d-4a6a-9808-5915f17ced30
Filesystem total blocks: 599854 (of 4096-byte blocks)
Filesystem total inodes: 2255795
Filesystem total metadata blocks: 74253
Filesystem total deduplicated bytes (of source files): 29625028195

user	2:35:08.03
system	1:12.65
total	2:39:25.35

$ ll -h perl-install-lzma9.erofs
-rw-r--r-- 1 mhx mhx 2.3G Apr 15 16:23 perl-install-lzma9.erofs

That’s definitely slower than SquashFS, but also significantly smaller.

For a fair comparison, let’s use the same 1 MiB block size with DwarFS,
but also tweak the options for best compression:

$ time mkdwarfs -i perl-install -o perl-install-1M.dwarfs -l9 -S20 -B64 --order=nilsimsa:max-cluster-size=150000
[...]
330733 dirs, 0/2440 soft/hard links, 1927501/1927501 files, 0 other
original size: 47.49 GiB, hashed: 43.47 GiB (1920025 files, 1.451 GiB/s)
scanned: 19.45 GiB (144675 files, 159.3 MiB/s), categorizing: 0 B/s
saved by deduplication: 28.03 GiB (1780386 files), saved by segmenting: 15.4 GiB
filesystem: 4.053 GiB in 4151 blocks (937069 chunks, 144674/144674 fragments, 144675 inodes)
compressed filesystem: 4151 blocks/806.2 MiB written
[...]
user	24:27.47
system	4:20.74
total	3:26.79

That’s significantly smaller and, almost more importantly, 46 times
faster than mkfs.erofs.

Actually using the file system images, here’s how DwarFS performs:

$ dwarfs perl-install-1M.dwarfs mnt -oworkers=8
$ find mnt -type f -print0 | xargs -0 -P16 -n64 cat | dd of=/dev/null bs=1M status=progress
50392172594 bytes (50 GB, 47 GiB) copied, 19 s, 2.7 GB/s
0+1662649 records in
0+1662649 records out
51161953159 bytes (51 GB, 48 GiB) copied, 19.4813 s, 2.6 GB/s

Reading every single file from 16 parallel processes took less than
20 seconds. The FUSE driver consumed 143 seconds of CPU time.

Here’s the same for EROFS:

$ erofsfuse perl-install-lzma9.erofs mnt
$ find mnt -type f -print0 | xargs -0 -P16 -n64 cat | dd of=/dev/null bs=1M status=progress
2594306810 bytes (2.6 GB, 2.4 GiB) copied, 300 s, 8.6 MB/s^C
0+133296 records in
0+133296 records out
2595212832 bytes (2.6 GB, 2.4 GiB) copied, 300.336 s, 8.6 MB/s

Note that I’ve stopped this after 5 minutes. The DwarFS FUSE driver
delivered about 300 times faster throughput compared to EROFS. The
EROFS FUSE driver consumed 50 minutes (!) of CPU time for only about
5% of the data, i.e. more than 400 times the CPU time consumed by
the DwarFS FUSE driver.

I’ve tried two more EROFS configurations on the same set of data.
The first one uses more or less just the defaults:

$ time mkfs.erofs -zlz4hc,12 perl-install-lz4hc.erofs perl-install
mkfs.erofs 1.7.1-gd93a18c9
Build completed.
------
Filesystem UUID: b75142ed-6cf3-46a4-84f3-12693f7759a0
Filesystem total blocks: 5847130 (of 4096-byte blocks)
Filesystem total inodes: 2255794
Filesystem total metadata blocks: 419699
Filesystem total deduplicated bytes (of source files): 0

user	3:38:23.36
system	1:10.84
total	3:41:37.33

The second one additionally enables the -Ededupe option:

$ time mkfs.erofs -zlz4hc,12 -Ededupe perl-install-lz4hc-dedupe.erofs perl-install
mkfs.erofs 1.7.1-gd93a18c9
Build completed.
------
Filesystem UUID: 0ccf581e-ad3b-4d08-8b10-5b7e15f8e3cd
Filesystem total blocks: 1510091 (of 4096-byte blocks)
Filesystem total inodes: 2255794
Filesystem total metadata blocks: 435599
Filesystem total deduplicated bytes (of source files): 19220717568

user	4:19:57.61
system	1:21.62
total	4:23:55.85

I don’t know why these are even slower than the first, seemingly more
complex, set of options. As was to be expected, the resulting images
were significantly bigger:

$ ll -h perl-install*.erofs
-rw-r--r-- 1 mhx mhx 5.8G Apr 16 02:46 perl-install-lz4hc-dedupe.erofs
-rw-r--r-- 1 mhx mhx  23G Apr 15 22:34 perl-install-lz4hc.erofs
-rw-r--r-- 1 mhx mhx 2.3G Apr 15 16:23 perl-install-lzma9.erofs

The good news is that these perform much better and even outperform
DwarFS, albeit by a small margin:

$ erofsfuse perl-install-lz4hc.erofs mnt
$ find mnt -type f -print0 | xargs -0 -P16 -n64 cat | dd of=/dev/null bs=1M status=progress
49920168315 bytes (50 GB, 46 GiB) copied, 16 s, 3.1 GB/s
0+1493031 records in
0+1493031 records out
51161953159 bytes (51 GB, 48 GiB) copied, 16.4329 s, 3.1 GB/s

The deduplicated version is even a tiny bit faster:

$ erofsfuse perl-install-lz4hc-dedupe.erofs mnt
find mnt -type f -print0 | xargs -0 -P16 -n64 cat | dd of=/dev/null bs=1M status=progress
50808037121 bytes (51 GB, 47 GiB) copied, 16 s, 3.2 GB/s
0+1499949 records in
0+1499949 records out
51161953159 bytes (51 GB, 48 GiB) copied, 16.1184 s, 3.2 GB/s

The EROFS kernel driver wasn’t any faster than the FUSE driver.

The FUSE driver used about 27 seconds of CPU time in both cases,
substantially less than before and 5 times less than DwarFS.

DwarFS can get close to the throughput of EROFS by using zstd instead
of lzma compression:

$ dwarfs perl-install-1M-zstd.dwarfs mnt -oworkers=8
find mnt -type f -print0 | xargs -0 -P16 -n64 cat | dd of=/dev/null bs=1M status=progress
49224202357 bytes (49 GB, 46 GiB) copied, 16 s, 3.1 GB/s
0+1529018 records in
0+1529018 records out
51161953159 bytes (51 GB, 48 GiB) copied, 16.6716 s, 3.1 GB/s

With fuse-archive

I came across fuse-archive
while looking for FUSE drivers to mount archives and it seems to be
the most versatile of the alternatives (and the one that actually
compiles out of the box).

An interesting test case straight from fuse-archive’s README is in
the Performance
section: an archive with a single huge file full of zeroes. Let’s
make the example a bit more extreme and use a 1 GiB file instead of
just 256 MiB:

$ mkdir zerotest
$ truncate --size=1G zerotest/zeroes

Now, we build several different archives and a DwarFS image:

$ time mkdwarfs -i zerotest -o zerotest.dwarfs -W16 --log-level=warn --progress=none

real    0m7.604s
user    0m7.521s
sys     0m0.083s

$ time zip -9 zerotest.zip zerotest/zeroes
  adding: zerotest/zeroes (deflated 100%)

real    0m4.923s
user    0m4.840s
sys     0m0.080s

$ time 7z a -bb0 -bd zerotest.7z zerotest/zeroes

7-Zip [64] 16.02 : Copyright (c) 1999-2016 Igor Pavlov : 2016-05-21
p7zip Version 16.02 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,64 bits,16 CPUs Intel(R) Xeon(R) E-2286M  CPU @ 2.40GHz (906ED),ASM,AES-NI)

Scanning the drive:
1 file, 1073741824 bytes (1024 MiB)

Creating archive: zerotest.7z

Items to compress: 1

Files read from disk: 1
Archive size: 157819 bytes (155 KiB)
Everything is Ok

real    0m5.535s
user    0m48.281s
sys     0m1.116s

$ time tar --zstd -cf zerotest.tar.zstd zerotest/zeroes

real    0m0.449s
user    0m0.510s
sys     0m0.610s

Turns out that tar --zstd is easily winning the compression speed
test. Looking at the file sizes did actually blow my mind just a bit:

$ ll zerotest.* --sort=size
-rw-r--r-- 1 mhx users 1042231 Jul  1 15:24 zerotest.zip
-rw-r--r-- 1 mhx users  157819 Jul  1 15:26 zerotest.7z
-rw-r--r-- 1 mhx users   33762 Jul  1 15:28 zerotest.tar.zstd
-rw-r--r-- 1 mhx users     848 Jul  1 15:23 zerotest.dwarfs

I definitely didn’t expect the DwarFS image to be that small.
Dropping the section index would actually save another 100 bytes.
So, if you want to archive lots of zeroes, DwarFS is your friend.

Anyway, let’s look at how fast and efficiently the zeroes can
be read from the different archives. First, the zip archive:

$ time dd if=mnt/zerotest/zeroes of=/dev/null status=progress
1020117504 bytes (1.0 GB, 973 MiB) copied, 2 s, 510 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 2.10309 s, 511 MB/s

real    0m2.104s
user    0m0.264s
sys     0m0.486s

CPU time used by the FUSE driver was 1.8 seconds and mount time
was in the milliseconds.

Now, the 7z archive:

 $ time dd if=mnt/zerotest/zeroes of=/dev/null status=progress
594759168 bytes (595 MB, 567 MiB) copied, 1 s, 595 MB/s
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 1.76904 s, 607 MB/s

real    0m1.772s
user    0m0.229s
sys     0m0.572s

CPU time used by the FUSE driver was 2.9 seconds and mount time
was just over 1.0 seconds.

Now, the .tar.zstd archive:

$ time dd if=mnt/zerotest/zeroes of=/dev/null status=progress
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 0.799409 s, 1.3 GB/s

real    0m0.801s
user    0m0.262s
sys     0m0.537s

CPU time used by the FUSE driver was 0.53 seconds and mount time
was 0.13 seconds.

Last but not least, let’s look at DwarFS:

$ time dd if=mnt/zeroes of=/dev/null status=progress
2097152+0 records in
2097152+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 0.753 s, 1.4 GB/s

real    0m0.757s
user    0m0.220s
sys     0m0.534s

CPU time used by the FUSE driver was 0.17 seconds and mount time
was less than a millisecond.

If we increase the block size for the dd command, we can get
even higher throughput. For fuse-archive with the .tar.zstd:

$ time dd if=mnt/zerotest/zeroes of=/dev/null status=progress bs=16384
65536+0 records in
65536+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 0.318682 s, 3.4 GB/s

real    0m0.323s
user    0m0.005s
sys     0m0.154s

And for DwarFS:

$ time dd if=mnt/zeroes of=/dev/null status=progress bs=16384
65536+0 records in
65536+0 records out
1073741824 bytes (1.1 GB, 1.0 GiB) copied, 0.172226 s, 6.2 GB/s

real    0m0.176s
user    0m0.020s
sys     0m0.141s

This is all nice, but what about a more real-life use case?
Let’s take the 1.82.0 boost release archives:

$ ll --sort=size boost_1_82_0.*
-rw-r--r-- 1 mhx users 208188085 Apr 10 14:25 boost_1_82_0.zip
-rw-r--r-- 1 mhx users 142580547 Apr 10 14:23 boost_1_82_0.tar.gz
-rw-r--r-- 1 mhx users 121325129 Apr 10 14:23 boost_1_82_0.tar.bz2
-rw-r--r-- 1 mhx users 105901369 Jun 28 12:47 boost_1_82_0.dwarfs
-rw-r--r-- 1 mhx users 103710551 Apr 10 14:25 boost_1_82_0.7z

Here are the timings for mounting each archive and then using
tar to build another archive from the mountpoint and just counting
the number of bytes in that archive, e.g.:

$ time tar cf - mnt | wc -c
803614720

real    0m4.602s
user    0m0.156s
sys     0m1.123s

Here are the results in terms of wallclock time and FUSE driver
CPU time:

Archive Mount Time tar Wallclock Time FUSE Driver CPU Time
.zip 0.458s 5.073s 4.418s
.tar.gz 1.391s 3.483s 3.943s
.tar.bz2 15.663s 17.942s 32.040s
.7z 0.321s 32.554s 31.625s
.dwarfs 0.013s 2.974s 1.984s

DwarFS easily wins all categories while still compressing the data
almost as well as 7z.

What about accessing files more randomly?

$ find mnt -type f -print0 | xargs -0 -P32 -n32 cat | dd of=/dev/null status=progress

It turns out that fuse-archive grinds to a halt in this case, so I had
to run the test on a subset (the boost subdirectory) of the data.
The .tar.bz2 and .7z archives were so slow to read that I stopped
them after a few minutes.

Archive Throughput Wallclock Time FUSE Driver CPU Time
.zip 1.8 MB/s 83.245s 83.669s
.tar.gz 1.2 MB/s 121.377s 122.711s
.tar.bz2 0.2 MB/s - -
.7z 0.3 MB/s - -
.dwarfs 598.0 MB/s 0.249s 1.099s

Performance Monitoring

Both the FUSE driver and dwarfsextract by default have support for
simple performance monitoring. You can build binaries without this
feature (-DENABLE_PERFMON=OFF), but impact should be negligible even
if performance monitoring is enabled at run-time.

To enable the performance monitor, you pass a list of components for which
you want to collect latency metrics, e.g.:

$ dwarfs test.dwarfs mnt -f -operfmon=fuse

When the driver exits, you will see output like this:

[fuse.op_read]
      samples: 45145
      overall: 3.214s
  avg latency: 71.2us
  p50 latency: 131.1us
  p90 latency: 131.1us
  p99 latency: 262.1us

[fuse.op_readdir]
      samples: 2
      overall: 51.31ms
  avg latency: 25.65ms
  p50 latency: 32.77us
  p90 latency: 67.11ms
  p99 latency: 67.11ms

[fuse.op_lookup]
      samples: 16
      overall: 19.98ms
  avg latency: 1.249ms
  p50 latency: 2.097ms
  p90 latency: 4.194ms
  p99 latency: 4.194ms

[fuse.op_init]
      samples: 1
      overall: 199.4us
  avg latency: 199.4us
  p50 latency: 262.1us
  p90 latency: 262.1us
  p99 latency: 262.1us

[fuse.op_open]
      samples: 16
      overall: 122.2us
  avg latency: 7.641us
  p50 latency: 4.096us
  p90 latency: 32.77us
  p99 latency: 32.77us

[fuse.op_getattr]
      samples: 1
      overall: 5.786us
  avg latency: 5.786us
  p50 latency: 8.192us
  p90 latency: 8.192us
  p99 latency: 8.192us

The metrics should be self-explanatory. However, note that the
percentile metrics are logarithmically quantized in order to use
as little resources as possible. As a result, you will only see
values that look an awful lot like powers of two.

Currently, the supported components are fuse for the FUSE
operations, filesystem_v2 for the DwarFS file system component
and inode_reader_v2 for the component that handles all read()
system calls.

The FUSE driver also exposes the performance monitor metrics via
an extended attribute.

Other Obscure Features

Setting Worker Thread CPU Affinity

This only works on Linux and usually only makes sense if you have CPUs
with different types of cores (e.g. “performance” vs “efficiency” cores)
and are really trying to squeeze the last ounce of speed out of DwarFS.

By setting the environment variable DWARFS_WORKER_GROUP_AFFINITY, you
can set the CPU affinity of different worker thread groups, e.g.:

export DWARFS_WORKER_GROUP_AFFINITY=blockify=3:compress=6,7

This will set the affinity of the blockify worker group to CPU 3 and
the affinity of the compress worker group to CPUs 6 and 7.

You can use this feature for all tools that use one or more worker thread
groups. For example, the FUSE driver dwarfs and dwarfsextract use a
worker group blkcache that the block cache (i.e. block decompression and
lookup) runs on. mkdwarfs uses a whole array of different worker groups,
namely compress for compression, scanner for scanning, ordering for
input ordering, and blockify for segmenting. blockify is what you would
typically want to run on your “performance” cores.

Stargazers over Time

Stargazers over Time