Author: Richard Plangger <[email protected]>
Branch: extradoc
Changeset: r5729:4401d9b4f722
Date: 2016-10-04 10:46 +0200
http://bitbucket.org/pypy/extradoc/changeset/4401d9b4f722/
Log: parts of the slides, not the final version
diff --git a/talk/pyconza2016/pypy/img/jitlog.png
b/talk/pyconza2016/pypy/img/jitlog.png
new file mode 100644
index
e69de29bb2d1d6434b8b29ae775ad8c2e48c5391..a1b9612df1d894ac8f8bf62045bb3e55307cb6f9
GIT binary patch
[cut]
diff --git a/talk/pyconza2016/pypy/img/vmprof-screen-pypy.png
b/talk/pyconza2016/pypy/img/vmprof-screen-pypy.png
new file mode 100644
index
e69de29bb2d1d6434b8b29ae775ad8c2e48c5391..55feb25897f117a07ee3edc644bfcfa213c4e739
GIT binary patch
[cut]
diff --git a/talk/pyconza2016/pypy/img/vmprof-screen.png
b/talk/pyconza2016/pypy/img/vmprof-screen.png
new file mode 100644
index
e69de29bb2d1d6434b8b29ae775ad8c2e48c5391..49d9fce98781b28ce0b1dffd906079e32b6cf6a9
GIT binary patch
[cut]
diff --git a/talk/pyconza2016/pypy/index.html b/talk/pyconza2016/pypy/index.html
--- a/talk/pyconza2016/pypy/index.html
+++ b/talk/pyconza2016/pypy/index.html
@@ -4,7 +4,7 @@
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0,
maximum-scale=1.0, user-scalable=no">
- <title>Compiler Workshop Austin Texas</title>
+ <title></title>
<link rel="stylesheet" href="css/reveal.css">
<link rel="stylesheet" href="css/theme/white.css">
@@ -32,13 +32,220 @@
</section>
<section>
<section>
- <h1>About me</h1>
- TODO
+ <h1>PyPy is ..</h1>
+ </section>
+ <section>
+ <p>... a software project ... </p>
+ </section>
+ <section>
+ <p>... assembling a <strong>fast virtual machine for
Python</strong> ... </p>
+ </section>
+ <section>
+ <p>developed by many researchers, freelancers and many
contributors.</p>
</section>
</section>
<section>
+ <section>
+ <h1>PyPy is not just that</h1>
+ </section>
+ <section>
+ <p>Experiment with new ideas</p>
+ </section>
+ <section>
+ <ul>
+ <li>Python written in Python</li>
+ <li>RPython</li>
+ <li>JIT compiler</li>
+ <li>VMProf</li>
+ <li>PyPy STM</li>
+ <li>...</li>
+ </ul>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>About me</h1>
+ <p>Working on PyPy (+1,5y)</p>
+ <p>Master degree - Sticked with PyPy</p>
+ <p>living and working in Austria</p>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>Speedy Python programs?</h1>
+ <p>When is your Python program fast enough?</p>
+ </section>
+ <section>
+ <p>When it gets a speeding ticket because it is too
fast?</p>
+ </section>
+ <section>
+ <p>or when PyPy's benchmark suite reaches 10x faster
on average?</p>
+ </section>
+ <section>
+ <p><strong>Neither</strong></p>
+ </section>
+ <section>
+ <p>Run you program an measure your criteria</p>
+ </section>
+ <section>
+ <h1>Criteria examples?</h1>
+ <ul>
+ <li>CPU time</li>
+ <li>Peak Heap Memory</li>
+ <li>Requests per second</li>
+ <li>...</li>
+ </ul>
+ <p>Dissatisfaction with one attribute of your
program!</p>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>Some theory ... </h1>
+ </section>
+ <section>
+ <h1>Hot spots</h1>
+ <p>Loops!</p>
+ <p>What kind program can you build without
loops?</p>
+ </section>
+ <section>
+ <h1>Complexity</h1>
+ <p>Big-O-Notation - Express how many steps a
program to complete at most</p>
+ </section>
+ <section>
+ <ul>
+ <li><code>a = 3</code> # runs in
O(1)</li>
+ <li><code>[x+1 for x in
range(n)]</code> # runs in O(n)</li>
+ <li><code>[[x+y for x in range(n)] for
y in range(m)]</code> # O(n*m)</li>
+ </ul>
+ </section>
+ <section>
+ <h1>Complexity</h1>
+ <p>Yields the most gain, independent from the
language</p>
+ <p>E.g. prefer O(n) over O(n**2)</p>
+ </section>
+ <section>
+ <h1>Only optimize a routine if ... </h1>
+ <p>you know that the complexity cannot be
stripped down</p>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>Let's start from the beginning</h1>
+ <p>with a small example</p>
+ </section>
+
+ <section>
+ <h1>Reading log files!</h1>
+ <p>JITLOG (facility to observe PyPy's JIT
internals)</p>
+ </section>
+
+ <section>
+ <ul>
+ <li>Written in Python</li>
+ <li>Moved to vmprof.com</li>
+ <li>Log files can easily take up to
40MB uncompressed</li>
+ <li>Takes ~14 seconds to parse with
CPython</li>
+ <li>Complexity is linear to input size
of the log file</li>
+ </ul>
+ </section>
+ <section>
+ <p class="advantage">+ Little development
time</p>
+ <p class="advantage">+ Easy to test</p>
+ <p><h3>Thanks to Python</h3></p>
+ </section>
+ <section>
+ <p class="disadvantage">- Takes too long to
parse</p>
+ <p>Our criteria: CPU time to long</p>
+ </section>
+ <section>
+ <p class="">Several possible ways</p>
+ <p>Caching</p>
+ <p>Reduce CPU time</p>
+ <p>Let's have both</p>
+ </section>
+ <section>
+ <p>Caching - Easily done with django caching
frame work</p>
+ <p>Reduce CPU time - Look at vmprof</p>
+ </section>
+ <section>
+ <h1>VMProf</h1>
+ <p><code>$ pip install vmprof</code></p>
+ <p><code>$ python -m vmprof --web
parse.py</code></p>
+ <p><a
href="http://vmprof.com/#/709eeba4d6bb3b52c19d8753fd65af7b";>→ link</a></p>
+ </section>
+ <section data-background="img/vmprof-screen.png">
+ </section>
+ <section data-background="img/vmprof-screen-pypy.png">
+ </section>
+ <section>
+ <h1>~4 times faster on PyPy</h1>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>Introducing PyPy's JIT</h1>
+ </section>
+ <section>
+ <h1>A simplified view</h1>
+ <ol>
+ <li>Start interpretation</li>
+ <li>Loops trigger recording</li>
+ <li>Optimization stage</li>
+ <li>Machine code generation</li>
+ </ol>
+ <p>Cannot represent control flow as a graph
(other than loop jumps)</p>
+ <p>Guards ensure correctness</p>
+ </section>
+ <section>
+ <h1>Beyond the scope of loops</h1>
+ <p>Frequent guard failure trigger recording</p>
+ </section>
+ <section data-background-image="img/jitlog.png">
+ <a
href="http://vmprof.com/#/7930e1f54f9eee75084738aafa6cb612/traces";>→
link</a>
+ </section>
+ <section>
+ <h1>JitViewer</h1>
+ <p>Tool to inspect PyPy internals</p>
+ <p>Helps you to learn and understand PyPy</p>
+ <p>Provided at vmprof.com</p>
+ </section>
+ </section>
+ <section>
+ <section>
+ <h1>Another real world example</h1>
+ <p></p>
+
+ </section>
+ <section>
+ <h1>Magnetic</h1>
+ <p>marketing tech company</p>
+ <p>switched to PyPy 3 years ago</p>
+ </section>
+ <section>
+ <h1>Q: what does your service do?</h1>
+ <p>A: ... allow generally large companies to
send targeted marketing (e.g. serve ads) to people based on data we have
learned </p>
+ </section>
+ <section>
+ <h1>Q: PyPy, where was it most helpful?</h1>
+ <p>A: ... ~30% speedups immediately from
switching to PyPy ... </p>
+ </section>
+ <section>
+ <h1>Q: PyPy issues?</h1>
+ <p>A: ... we had to solve for rolling deploys
... but that's ok, that's fairly easy ... </p>
+ </section>
+ <section>
+ <h1>Q: Value to your company?</h1>
+ <p>A: Latency speedup was somewhere aroudn 10%
...</p>
+ <p>But that number is deceiving</p>
+ <p>It's very valuable for us obviously</p>
+ <p>But it's only 10%, because even this app
that I'm talking about, which is fairly high volume (500,000 QPS), is a WSGI
app</p>
+ <p>So it spends lots of time blocking</p>
+ </section>
+ </section>
+ <section>
<h4>Questions?</h4>
- <a href="morepypy.blogspot.com">morepypy.blogspot.com</a>
+ <a
href="morepypy.blogspot.com">morepypy.blogspot.com</a><br>
+ Join on IRC <a href="">#pypy</a>
</section>
</div>
</div>
_______________________________________________
pypy-commit mailing list
[email protected]
https://mail.python.org/mailman/listinfo/pypy-commit
