[pypy-svn] r77934 - in pypy/extradoc/talk/pepm2011: . benchmarks
cfbolz at codespeak.net
cfbolz at codespeak.net
Thu Oct 14 15:37:30 CEST 2010
Author: cfbolz
Date: Thu Oct 14 15:37:28 2010
New Revision: 77934
Modified:
pypy/extradoc/talk/pepm2011/benchmarks/bench.gnumeric
pypy/extradoc/talk/pepm2011/escape-tracing.pdf
pypy/extradoc/talk/pepm2011/paper.tex
Log:
(cfbolz, arigo for some of it):
- add URLs for PyPy
- add factors to the timing table, to read it more easily
- change psyco description
- some fixes in the virtualizables section
Modified: pypy/extradoc/talk/pepm2011/benchmarks/bench.gnumeric
==============================================================================
Binary files. No diff available.
Modified: pypy/extradoc/talk/pepm2011/escape-tracing.pdf
==============================================================================
Binary files. No diff available.
Modified: pypy/extradoc/talk/pepm2011/paper.tex
==============================================================================
--- pypy/extradoc/talk/pepm2011/paper.tex (original)
+++ pypy/extradoc/talk/pepm2011/paper.tex Thu Oct 14 15:37:28 2010
@@ -211,7 +211,7 @@
\subsection{PyPy}
\label{sub:PyPy}
-The work described in this paper was done in the context of the PyPy project
+The work described in this paper was done in the context of the PyPy project\footnote{\texttt{http://pypy.org}}
\cite{armin_rigo_pypys_2006}. PyPy is an environment where dynamic languages can
be implemented in a simple yet efficient way.
When implementing a language with PyPy one writes an \emph{interpreter}
@@ -933,7 +933,7 @@
frame-introspection features in many dynamic languages. Languages such
as Python and Smalltalk allow the programmer to get access to the
frames object that the interpreter uses to store local variables. This
-is a useful feature, as makes the implementation of a debugger
+is a useful feature, as it makes the implementation of a debugger
possible in Python without needing much explicit support from the VM
level. On the other hand, it severely hinders the effectiveness of
allocation removal, because every time an object is stored into a
@@ -947,12 +947,14 @@
because a debugger is used) using values from JIT-emitted code runtime
locations (typically the CPU stack). Therefore in the common case,
nothing is stored into the frame objects, making the problem of too
-much escaping go away. This is a common approach in VM implementations
-\cite{miranda_context_1999}, the only novelty in our approach lays in
-its generality, all the delicate support code for this is generated, because
-most other JITs are just specifically written for one particular
+much escaping go away. This special handling of frame objects is a common approach in VM implementations
+\cite{miranda_context_1999}. The only novelty in our approach lays in
+its generality: all the delicate support code for this is generated, as opposed to
+most other JITs, which are just specifically written for one particular
language.
+XXX one of the JS tracing JITs does this as well
+
% subsection Virtualizables (end)
% section Supporting Techniques (end)
@@ -963,7 +965,10 @@
To evaluate the effectiveness of our allocation removal algorithm, we look at
the effectiveness when used in the tracing JIT of PyPy's Python interpreter. The
benchmarks we used are small-to-medium Python programs, some synthetic
-benchmarks, some real applications.
+benchmarks, some real applications.\footnote{All the source code of the
+benchmarks can be found at \texttt{http://codespeak.net/svn/pypy/benchmarks/}.
+There is also a website that monitors PyPy's performance nightly at
+\texttt{http://speed.pypy.org/}.}
XXX [fijal] note that some of those have cpython-specific hacks removed or
old versions used (nbody mostly I think).
@@ -991,6 +996,7 @@
framework\footnote{\texttt{http://twistedmatrix.com/}}.
\end{itemize}
+
We evaluate the allocation removal algorithm along two lines: first we want to
know how many allocations could be optimized away. On the other hand, we want
to know how much the run times of the benchmarks is improved.
@@ -1025,7 +1031,7 @@
\begin{center}
\begin{tabular}{|l||r|rr|rr|rr|rr|}
\hline
- &num loops &new &removed &get/set &removed &guard &removed &all ops &removed\\
+ &num loops &new &removed &get/set &removed &guard &removed &all ops &removed\\
\hline
crypto\_pyaes &78 &3088 &50\% &57148 &25\% &9055 &95\% &137189 &80\%\\
django &51 &673 &54\% &19318 &18\% &3876 &93\% &55682 &85\%\\
@@ -1061,32 +1067,30 @@
All benchmarks are improved by the allocation removal optimization, some as much
as XXX. XXX Psyco
-XXX runtimes of the algorithm somehow?
-
\begin{figure*}
\begin{center}
-\begin{tabular}{|l||r|r|r|r|}
+\begin{tabular}{|l||r|r||r|r||r|r||r|r|}
\hline
- &CPython &Psyco &PyPy w/o optimizations &PyPy w/ optimizations \\
+ &CPython [ms] & $\times$ &Psyco [ms] & $\times$ &PyPy w/o optim. [ms] & $\times$ &PyPy w/ optim. [ms]& $\times$ \\
\hline
-crypto\_pyaes &2757.80 $\pm$ 0.98 &67.90 $\pm$ 0.47 &1652.00 $\pm$ 4.00 &266.86 $\pm$ 5.94 \\
-django &993.19 $\pm$ 0.50 &913.51 $\pm$ 4.22 &694.73 $\pm$ 2.86 &259.53 $\pm$ 1.79 \\
-fannkuch &1987.22 $\pm$ 2.02 &944.44 $\pm$ 0.61 &566.99 $\pm$ 1.06 &466.87 $\pm$ 1.85 \\
-go &947.21 $\pm$ 1.58 &445.96 $\pm$ 0.68 &2197.71 $\pm$ 25.21 &316.15 $\pm$ 9.33 \\
-html5lib &13987.12 $\pm$ 19.51 &17398.25 $\pm$ 36.50 &27194.45 $\pm$ 46.62 &10092.19 $\pm$ 23.50 \\
-meteor-contest &346.98 $\pm$ 0.35 &215.66 $\pm$ 0.23 &433.04 $\pm$ 1.45 &392.85 $\pm$ 0.87 \\
-nbody\_modified &637.90 $\pm$ 1.82 &256.78 $\pm$ 0.18 &135.55 $\pm$ 0.33 &103.93 $\pm$ 0.25 \\
-pyflate-fast &3169.35 $\pm$ 1.89 &1278.16 $\pm$ 3.13 &3285.89 $\pm$ 8.51 &1822.36 $\pm$ 11.52 \\
-raytrace-simple &2744.60 $\pm$ 51.72 &1072.66 $\pm$ 1.08 &2778.27 $\pm$ 15.13 &647.24 $\pm$ 5.44 \\
-richards &354.06 $\pm$ 1.00 &63.48 $\pm$ 0.15 &383.93 $\pm$ 3.28 &88.32 $\pm$ 0.91 \\
-spambayes &299.16 $\pm$ 0.35 &338.68 $\pm$ 3.14 &580.90 $\pm$ 24.68 &397.37 $\pm$ 10.60 \\
-spectral-norm &478.63 $\pm$ 0.80 &139.83 $\pm$ 1.54 &353.51 $\pm$ 1.39 &112.10 $\pm$ 1.17 \\
-telco &1207.67 $\pm$ 2.03 &730.00 $\pm$ 2.66 &1296.08 $\pm$ 4.37 &495.23 $\pm$ 2.14 \\
-twisted\_names &9.58 $\pm$ 0.01 &10.43 $\pm$ 0.01 &17.99 $\pm$ 0.27 &7.13 $\pm$ 0.09 \\
+crypto\_pyaes &2757.80 $\pm$ 0.98 &10.33 &67.90 $\pm$ 0.47 &0.25 &1652.00 $\pm$ 4.00 &6.19 &266.86 $\pm$ 5.94 &1.00\\
+django &993.19 $\pm$ 0.50 &3.83 &913.51 $\pm$ 4.22 &3.52 &694.73 $\pm$ 2.86 &2.68 &259.53 $\pm$ 1.79 &1.00\\
+fannkuch &1987.22 $\pm$ 2.02 &4.26 &944.44 $\pm$ 0.61 &2.02 &566.99 $\pm$ 1.06 &1.21 &466.87 $\pm$ 1.85 &1.00\\
+go &947.21 $\pm$ 1.58 &3.00 &445.96 $\pm$ 0.68 &1.41 &2197.71 $\pm$ 25.21 &6.95 &316.15 $\pm$ 9.33 &1.00\\
+html5lib &13987.12 $\pm$ 19.51 &1.39 &17398.25 $\pm$ 36.50 &1.72 &27194.45 $\pm$ 46.62 &2.69 &10092.19 $\pm$ 23.50 &1.00\\
+meteor-contest &346.98 $\pm$ 0.35 &0.88 &215.66 $\pm$ 0.23 &0.55 &433.04 $\pm$ 1.45 &1.10 &392.85 $\pm$ 0.87 &1.00\\
+nbody\_modified &637.90 $\pm$ 1.82 &6.14 &256.78 $\pm$ 0.18 &2.47 &135.55 $\pm$ 0.33 &1.30 &103.93 $\pm$ 0.25 &1.00\\
+pyflate-fast &3169.35 $\pm$ 1.89 &1.74 &1278.16 $\pm$ 3.13 &0.70 &3285.89 $\pm$ 8.51 &1.80 &1822.36 $\pm$ 11.52 &1.00\\
+raytrace-simple &2744.60 $\pm$ 51.72 &4.24 &1072.66 $\pm$ 1.08 &1.66 &2778.27 $\pm$ 15.13 &4.29 &647.24 $\pm$ 5.44 &1.00\\
+richards &354.06 $\pm$ 1.00 &4.01 &63.48 $\pm$ 0.15 &0.72 &383.93 $\pm$ 3.28 &4.35 &88.32 $\pm$ 0.91 &1.00\\
+spambayes &299.16 $\pm$ 0.35 &0.75 &338.68 $\pm$ 3.14 &0.85 &580.90 $\pm$ 24.68 &1.46 &397.37 $\pm$ 10.60 &1.00\\
+spectral-norm &478.63 $\pm$ 0.80 &4.27 &139.83 $\pm$ 1.54 &1.25 &353.51 $\pm$ 1.39 &3.15 &112.10 $\pm$ 1.17 &1.00\\
+telco &1207.67 $\pm$ 2.03 &2.44 &730.00 $\pm$ 2.66 &1.47 &1296.08 $\pm$ 4.37 &2.62 &495.23 $\pm$ 2.14 &1.00\\
+twisted\_names &9.58 $\pm$ 0.01 &1.34 &10.43 $\pm$ 0.01 &1.46 &17.99 $\pm$ 0.27 &2.52 &7.13 $\pm$ 0.09 &1.00\\
\hline
\end{tabular}
\end{center}
-\caption{Benchmark Times in Milliseconds}
+\caption{Benchmark Times in Milliseconds, Together With Factor Over PyPy With Optimizations}
\label{fig:times}
\end{figure*}
@@ -1121,10 +1125,11 @@
implementation are given.
Psyco \cite{rigo_representation-based_2004} is a (non-tracing) JIT for Python
-that fully implements allocation removal (the \emph{static} objects are called
-\emph{virtual} objects). It is a hand-written extension module for CPython.
-Historically, PyPy's JIT can be seen as some successor of Psyco for a
-general context.
+that implements a more ad-hoc version of the allocation removal described here.
+Our static objects could be related to what are called \emph{virtual} objects
+in Psyco. It is a hand-written extension module for CPython. Historically,
+PyPy's JIT can be seen as some successor of Psyco for a general context (one of
+the authors of this paper is the author of Psyco).
partial evaluation:
More information about the Pypy-commit
mailing list