<!DOCTYPE html>
<html lang="en">
<head>
<meta content="width=device-width, initial-scale=1" charset="utf-8" />
<title>Benchmarking and comparing DwarFS</title>
<link href="/style.css" type="text/css" rel="stylesheet" />
<link href="/prism.css" type="text/css" rel="stylesheet" />
</head>
<body class="line-numbers">
<h1 id="benchmarking-and-comparing-dwarfs">Benchmarking and
comparing DwarFS</h1>
<p>DwarFS is a filesystem developed by the user mhx on GitHub
(<em>mhx/dwarfs</em>), which is self-described as "A fast high
compression read-only file system for Linux, Windows, and
macOS." One of my ideas for blendOS was to layer different
packages, and combined with its compression and option to be
mounted as a FUSE-based filesystem, it's an appealing option for
this use case - blendOS is immutable, so it might as well have
some compression.</p>
<h2 id="methodology">Methodology</h2>
<p>The datasets being used for this test will be the
following:</p>
<ul>
<li>25 GiB of null data (just <code>00000000</code> in
binary)</li>
<li>25 GiB of random data<a href="#fn1" class="footnote-ref"
id="fnref1" role="doc-noteref"><sup>1</sup></a></li>
<li>Data for a 100 million-sided regular polygon; ~26.5 GiB<a
href="#fn2" class="footnote-ref" id="fnref2"
role="doc-noteref"><sup>2</sup></a></li>
<li>The current Linux longterm release source (<a
href="https://cdn.kernel.org/pub/linux/kernel/v6.x/linux-6.6.58.tar.xz">6.6.58</a>
(<em>The Linux Kernel Archives</em>)); ~1.5 GB</li>
<li>For some rough latency testing:
<ul>
<li>1024 4 KiB files filled with null data (again, just
<code>00000000</code> in binary)</li>
<li>1024 4 KiB files filled with random data</li>
</ul></li>
</ul>
<p>All this data should cover both latency and read speed
testing for data that compresses differently - extremely
compressible files with null data, decently compressible files,
and random data which can't be compressed well.</p>
<h3 id="what-filesystems">What filesystems?</h3>
<p>I'll be benchmarking DwarFS (<em>mhx/dwarfs</em>),
fuse-archive (<em>Google/Fuse-Archive</em>) (with tar files),
and btrfs. In some early, basic testing, I found that mounting
any <em>compressed</em> archives with <code>fuse-archive</code>,
a tool for mounting archive file formats as read-only
filesystems, took far too long. Additionally, being FUSE-based,
these would have slightly worse performance than kernel
filesystems, so I tried to use a FUSE driver as well for btrfs.
Unforunately, I ran into a bug, so I won't be able to quite do
an equivalent test; btrfs will only be running in the
kernel.</p>
<p>During said early testing, I also ran into the fact that most
compressed archives, like Gzip-compressed tar archives, also
took far too long to <em>create</em>, because Gzip is
single-threaded. So all the options with no chance of being used
have been marked off, and I'll only be looking into these
three.</p>
<p>DwarFS also took far too long to create an archive on its
default setting, but on compression level 1, it's much faster -
11m2.738s for the ~80 GiB total, and considering my entire
system is about 20 GiB, that should be about 2-3 minutes, which
is reasonable; With no compression, tar took 3m3.378s. Mounting
the DwarFS archive was nearly instant (0.022s), while mounting
the tar archive took 1.352s - not bad, but not ideal when
mounting many, and will absolutely be taken into
consideration.</p>
<h2 id="running-the-benchmark">Running the benchmark</h2>
<p>First off, installed I installed my benchamark (<em>Disk Read
Benchmark</em>) by cloning the repository, installing it using
Cargo, then added its completions to fish (just for this
session):</p>
<div class="sourceCode" id="cb2"><pre
class="language-sh"><code class="language-bash"><span id="cb2-1"><a href="#cb2-1" aria-hidden="true" tabindex="-1"></a><span class="fu">git</span> clone https://git.askiiart.net/askiiart/disk-read-benchmark</span>
<span id="cb2-2"><a href="#cb2-2" aria-hidden="true" tabindex="-1"></a><span class="bu">cd</span> ./disk-read-benchmark</span>
<span id="cb2-3"><a href="#cb2-3" aria-hidden="true" tabindex="-1"></a><span class="ex">cargo</span> install <span class="at">--path</span> .</span>
<span id="cb2-4"><a href="#cb2-4" aria-hidden="true" tabindex="-1"></a><span class="ex">disk-read-benchmark</span> generate-fish-completions <span class="kw">|</span> <span class="bu">source</span></span></code></pre></div>
<p>Then I prepared all the data:</p>
<div class="sourceCode" id="cb3"><pre
class="language-sh"><code class="language-bash"><span id="cb3-1"><a href="#cb3-1" aria-hidden="true" tabindex="-1"></a><span class="ex">disk-read-benchmark</span> prep-dirs</span>
<span id="cb3-2"><a href="#cb3-2" aria-hidden="true" tabindex="-1"></a><span class="ex">disk-read-benchmark</span> grab-data</span>
<span id="cb3-3"><a href="#cb3-3" aria-hidden="true" tabindex="-1"></a><span class="ex">./prepare.sh</span></span></code></pre></div>
<p><code>disk-read-benchmark</code> prepares all the
directories, generates the data to be used for testing, then
<code>./prepare.sh</code> uses the data to generate the DwarFS
and tar archives.</p>
<p>To run it, I just ran this:</p>
<div class="sourceCode" id="cb4"><pre
class="language-sh"><code class="language-bash"><span id="cb4-1"><a href="#cb4-1" aria-hidden="true" tabindex="-1"></a><span class="ex">disk-read-benchmark</span> benchmark</span></code></pre></div>
<p>Which outputs the data at
<code>data/benchmark-data.csv</code> and
<code>data/bulk.csv</code> for the single and bulk files,
respectively.</p>
<h2 id="results">Results</h2>
<p>After processing <a
href="/assets/benchmarking-dwarfs/data/">the data</a> with <a
href="/assets/benchmarking-dwarfs/process-data.py">this
script</a> to make it a bit easier, I put the resulting graphs
in here ↓</p>
<h3 id="sequential-read">Sequential read</h3>
<p>These results interest me quite a bit; unsurprisingly, DwarFS
has an advantage on the null file, due to its compression,
though it's disappointing the difference in time wasn't greater.
However, it does far worse on the random file, and I'm not sure
why; as discussed further down, DwarFS doesn't try to compress
incompressible files as far as I know, but I could be wrong. As
for the 100 million-sided polygon, it's somewhere in between,
with an advantage due to its compression, but still taking
longer than expected.</p>
<p>As for fuse-archive, it handles the null file well, but takes
longer on the others; not much to say.</p>
<div>
<canvas id="seq_read_chart" class="chart">
</canvas>
</div>
<h3 id="random-read">Random read</h3>
<p>There's nothing much to say here; although DwarFS took
significantly longer, it's still pretty fast - a different of
about 14 milliseconds worst case, across a 25 GiB file; similar
resuls for the 100 million-sided polygon, though to a less
extent, given it can be compressed better. With the null file,
due to its compression, DwarFS was actually on par with
fuse-archive, but it can't compete with btrfs's performance,
given it's so heavily optimized, and in the kernel.</p>
<div>
<canvas id="rand_read_chart" class="chart">
</canvas>
</div>
<h3 id="sequential-read-latency">Sequential read latency</h3>
<p>As expected, DwarFS performs a bit worse on the
incompressible random data, but otherwise they'll all roughly
equal. I wasn't expecting this, given btrfs is in the kernel,
while the other two are using FUSE.</p>
<div>
<canvas id="seq_read_latency_chart" class="chart">
</canvas>
</div>
<h3 id="random-read-latency">Random read latency</h3>
<p>Both DwarFS and fuse-archive had some trouble with this test.
DwarFS doesn't seem to handle random access very well; this is
supposedly fixed, as seen in issue 139 (<em>Issue #139 ·
mhx/dwarfs</em>), but the performance issues are obvious
regardless; I'm not sure why, given it doesn't compress
uncompressible data, not to mention it does just fine on the
random read test, where the only difference is that it reads
<em>more</em> data. But regardless, DwarFS ended up performing
far worse than expected on both the incompressible random data,
and the highly compressible null data.</p>
<p>Meanwhile, when testing random read latency in
<code>fuse-archive</code> pretty much just dies, becoming
ridiculously slow (even compared to DwarFS), so I didn't include
its single-file results. It succeeds on the bulk files, but
given it just shows as 0 seconds anyways, given the massive
scale, I opted to not include it in this graph at all.</p>
<div>
<canvas id="rand_read_latency_chart" class="chart">
</canvas>
</div>
<h2 id="misc-notes">Misc notes</h2>
<p>DwarFS can take up a fair amount of memory if mounting it
many times (<em>Issue #219 · mhx/dwarfs</em>), and this should
be kept in mind for use in BlendOS.</p>
<hr />
<p>Ratarmount (<em>mxmlnkn/ratarmount</em>) should also be
investigated; it's similar to fuse-archive, but with some
improvements, and some important notes. From its README
file:</p>
<blockquote>
<p>Note that fuse-archive daemonizes instantly but the mount
point will not be usable for a long time and everything trying
to use it will hang until then when not using
--asyncprogress</p>
</blockquote>
<blockquote>
<p>Mounting bzip2 and xz archives has actually become faster
than archivemount and fuse-archive with ratarmount -P 0 on most
modern processors because it actually uses more than one core
for decoding those compressions. indexed_bzip2 supports block
parallel decoding since version 1.2.0.</p>
</blockquote>
<p>Despite being written in Python, Ratarmount seems to have
significant performance improvements over fuse-archive.</p>
<hr />
<p>This should also be tested on systems with different specs,
like my Chromebook and laptop, and should try getting the btrfs
FUSE driver working and benchmarking that.</p>
<h2 id="summary">Summary</h2>
<p>DwarFS, or just the normal filesystem plus overlayfs, seem
like they may be the best options - DwarFS's compression and
deduplication are great, and the deduplication could probably be
used in way I haven't even thought of yet, but it has some niche
issues. Overall, I'm leaning towards using DwarFS as an option,
with just overlayfs as the default, but further testing is
needed.</p>
<h2 id="footnotes">Footnotes</h2>
<h2 id="sources">Sources</h2>
<p> - “Confused_ace_noises/Maths-Demos - Branch:
Headless-Deterministic.” Forgea: Git with a Cup of Jea,
git.askiiart.net/confused_ace_noises/maths-demos/src/branch/headless-deterministic.<br />
- “Disk Read Benchmark - A Simple and Performant Read-Only Disk
Benchmark, Written in Rust.” Forgea: Git with a Cup of Jea,
git.askiiart.net/askiiart/disk-read-benchmark.<br />
- Google. “Google/Fuse-Archive: Fuse File System for Archives
and Compressed Files (ZIP, RAR, 7z, ISO, TGZ, Xz...).” GitHub,
github.com/google/fuse-archive.<br />
- The Linux Kernel Archives, Linux Kernel Organization, Inc.,
<www.kernel.org/>.<br />
- Mhx. “Feature Request: Improve Block Management for
Uncompressed Blocks to Save Memory and Enhance Deduplication ·
ISSUE #139 · MHX/Dwarfs.” GitHub,
github.com/mhx/dwarfs/issues/139.<br />
- mhx. “mhx/Dwarfs: A Fast High Compression Read-Only File
System for Linux, Windows and Macos.” GitHub,
github.com/mhx/dwarfs.<br />
- mhx. “[Feature Request] Mounting Multiple Archives to the
Same Path · Issue #219 · MHX/Dwarfs.” GitHub,
github.com/mhx/dwarfs/issues/219.<br />
- mxmlnkn. “mxmlnkn/ratarmount: Access Large Archives as a
Filesystem Efficiently, e.g., Tar, Rar, Zip, Gz, BZ2, XZ, ZSTD
Archives.” GitHub, github.com/mxmlnkn/ratarmount.</p>
<!-- JavaScript for graphs goes hereeeeeee -->
<script src="https://cdn.jsdelivr.net/npm/chart.js"></script>
<script src="/assets/benchmarking-dwarfs/js/declare_vars.js"></script>
<script src="/assets/benchmarking-dwarfs/js/seq_read.js"></script>
<script src="/assets/benchmarking-dwarfs/js/rand_read.js"></script>
<script src="/assets/benchmarking-dwarfs/js/seq_latency.js"></script>
<script src="/assets/benchmarking-dwarfs/js/rand_latency.js"></script>
<section id="footnotes"
class="footnotes footnotes-end-of-document" role="doc-endnotes">
<hr />
<ol>
<li id="fn1"><p>My code can generate up to 25 GB/s. However, it
does random writes to my drive, which is <em>much</em> slower.
So on one hand, you could say my code is so amazingly fast that
current day technologies simply can't keep up. Or you could say
that I have no idea how to code for real world scenarios.<a
href="#fnref1" class="footnote-back"
role="doc-backlink">↩︎</a></p></li>
<li id="fn2">This data is from a modified version of an
abandoned math demonstration program
(<em>confused_ace_noises/maths-demos</em>) made by a friend; it
generates regular polygons and writes their data to a file. I
chose this because it was an artificial and reproducible yet
fairly compressible dataset (without being extremely
compressible like null data).
<details open>
<summary>
3-sided regular polygon data
</summary>
<br>
<!-- I put it in here just as a `style`, it didn't work. I put it in as a div with that `style`, it didn't work. I put it in as a div of that class which has those properties in style.css, it works -->
<!-- i hate webdev i hate webdev i hate webdev i hate webdev i hate webdev i hate webdev -->
<div class="force-word-wrap">
<pre><code>[Vertex { position: Pos([0.5, 0.0, 0.0]), color: Col([0.5310667, 0.7112941, 0.7138775]) }, Vertex { position: Pos([-0.25000003, 0.4330127, 0.0]), color: Col([0.7492257, 0.3142163, 0.49905664]) }, Vertex { position: Pos([0.0, 0.0, 0.0]), color: Col([0.2046682, 0.25598457, 0.72071356]) }, Vertex { position: Pos([-0.25000003, 0.4330127, 0.0]), color: Col([0.6389981, 0.5204368, 0.077735074]) }, Vertex { position: Pos([-0.24999996, -0.43301272, 0.0]), color: Col([0.8869035, 0.30709425, 0.8658899]) }, Vertex { position: Pos([0.0, 0.0, 0.0]), color: Col([0.2046682, 0.25598457, 0.72071356]) }, Vertex { position: Pos([-0.24999996, -0.43301272, 0.0]), color: Col([0.6236294, 0.03584433, 0.7590722]) }, Vertex { position: Pos([0.5, 8.742278e-8, 0.0]), color: Col([0.6105084, 0.3593351, 0.85544324]) }, Vertex { position: Pos([0.0, 0.0, 0.0]), color: Col([0.2046682, 0.25598457, 0.72071356]) }]</code></pre>
</div>
</details>
<a href="#fnref2" class="footnote-back"
role="doc-backlink">↩︎</a></li>
</ol>
</section>
<iframe src="https://john.citrons.xyz/embed?ref=askiiart.net" style="margin-left:auto;display:block;margin-right:auto;max-width:732px;width:100%;height:94px;border:none;"></iframe>
<script src="/prism.js"></script>
</body>
<footer>
<p><a href="https://git.askiiart.net/askiiart/engl-2311-blog">Source code</a> | <a href="/feed.xml">RSS</a> | <a href="/glossary.html">Glossary</a> | <a href="/about.html">About</a></p>
<small>Image captions are the same as the alt text; assuming you're sighted, you can most likely ignore them.</small>
</footer>
</html>