[pypy-dev] Hildesheim Coding Sprint Report
micahel at gmail.com
Wed Mar 7 15:12:02 CET 2007
This is the report for the second, coding, half of PyPy's **25th**
sprint, and the 18th and final sprint of the EU funded period.
We are so completely tired that we don't have the energy to write a
witty entry, so we'll skip that bit and start with describing the now
usual tutorial for those participants who are less familiar with PyPy
code base. This time Carl Friedrich was talking to Georg Brandl, an
interested CPython developer, and Anders (only the third Anders to be
involved in the project!) Sigfridsson, a PhD student at the
Interaction Design Centre at the University of Limerick. Anders
originally became involved in the project as an observer of the sprint
process for his group's research when we sprinted at the University of
Limerick, and that was partly why he was here this time, but it seems
he found PyPy interesting enough to learn Python in the mean time and
participate on the coding less at this sprint (or maybe he thought the
view from the sprint trenches would be more revealing!).
This sprint has seen many, many small tasks completed. On the first
morning, Holger and Armin improved the readline module to the point of
being useful -- supporting line editing and history, but not
completion -- and hooked it into the interpreter sufficiently that the
interactive interpreter and pdb both use it when available. At the
same time Richard and Michael were hunting a bug Richard had
discovered translating his own code, which was generally referred to
as "the rdict bug" but turned out to be a bug in the garbage
Carl Friedrich and his band of helpers (mostly Anders, Georg,
Alexander) worked on experimental reimplementations of Python lists,
one using a theoretically optimal overallocation strategy and another
using chunked storage to reduce the cost of list resizing. Sadly both
resulted in a measurable slow down. This can be seen as yet more
evidence that theory is different from practice...
CF's other target for reimplementation at this sprint was strings.
With help and moral support from Armin, he reimplemented strings
according to the design from the "ropes paper" of Boehm, Atkinson &
The predictable effect was that "typical Python code", written in the
knowledge of how strings are implemented in Python today, takes a
small (at most 10%) performance hit, but an arguably more natural and
naive style of string handling becomes efficient. And some completely
crazy code (like hash('a'*sys.maxint)) becomes very fast too...
Continuing the theme of making things slower by object
reimplementation, Armin supplied an implementation of general
dictionaries as a hash table whose collision resolution is via chained
lists instead of open addressing. Next!
As opposed to the above, Armin and CF implemented caching of app-level
character (i.e. strings of length 1) objects, which was a clear win,
improving the pystone benchmark by around 10%.
There have been many discussions recently about optimizing the lookup
of global variables, and during one that took place here about various
corner cases, Armin and Carl Friedrich and Samuele removed from PyPy
some of the strange things CPython does to determine what the
__builtins__ are for the execution of a given frame -- of course,
depending on the value of PyPy's five millionth configuration option.
Holger and Antonio came up with yet another optimization idea along
these lines, which can be found in doc/discussion/chained_getattr.txt.
Going back to the first day, Anto and Samuele worked on analyzing why
pypy-cli was being reported as 50+ times slower than CPython on the
To do this they wrote some small benchmarks in RPython and stared at
some code, but the main problem seems to be that Mono on PowerPC just
isn't that good: running pypy-cli using Microsoft's runtime shows
performance just 3-4 times slower than pypy-c.
After this, they worked on streamlining PyPy's much complained about
external function interface (and also broke translation a few times in
the process). The last sprint saw the introduction of a more general
registry-based interface for external functions, and Samuele and Anto
began by moving the math module over to using this interface. This
was harder than what had gone before because these functions depend on
header files, so some modifications to the C and LLVM backends were
On the last day, Anto made some small improvements to pypy-cli's
performance and Samuele made the taint object space translatable.
On the first day Georg and Alexander tried to see how fast a PyPy
could get if there was no Global Interpreter Lock (GIL). By disabling
the GIL and making the exception state thread-local on the genc-level,
they could easily get a binary that at least didn't crash if multiple
threads were not modifying internal stuff concurrently. Running 4
Pystone instances (from 4 different modules) on this pypy-c let the
process use 381% of cpu time, but the resulting figures were
disappointing: running the 4 Pystone instances in parallel was less
than 25% faster than running them in series, as opposed to being 300%
faster in the best case. Both concluded that the garbage collector
used (Boehm) is not very well suited for the heavy-duty memory
allocation pattern of PyPy in case of multiple threads.
After this, they implemented some of Python 2.5's features in the
interpreter, in particular support for __index__ and some extensions
to string and dict methods.
Anders and Anders worked very productively on fixing some of the bugs
in PyPy's issue tracker, implementing the -m command line option in
pypy-c, much improved handling of EINTR results from syscalls (which
makes most difference when pressing ^C on the command line), allowing
buffer objects to be passed to socket.send and preventing modifications
to builtin types.
Holger and Stephan worked in the direction of moving the currently
app-level and extremely slow string interpolation code into RPython by
separating out the code that analyzes the format string from the code
that access the values to be interpolated.
generally tidying and improving compatibility with Internet Explorer.
Guido should not be allowed to forget saying "I am happy to work with
Internet Explorer" during one of the daily status meetings :-)
Stephan and Arre worked on fixing the last remaining bugs in the
rewrite of the application-level stackless code that Stephan had been
working on for some time.
Later Stephan joined Armin and Christian in a discussion about the
best API to provide for the new "composable coroutine" concept. They
feel that the concept is powerful to encompass threads, greenlets,
coroutines, threads, syslets and the best way to barbecue ribs. You
can read about the basic idea in the "Composability" section of PyPy's
and further insight is unlikely to be provided by this diagram:
The basic conclusion was that this is a very nice and natural model
for a lot of things, at least once you've whacked your head into the
A task that occupied various people at various times of the sprint was
that of benchmarking, the goal being to determine how much effect the
object and other optimizations have. Michael had over the previous
month or so some written some scaffolding code to allow various
benchmarks to be run and the results recorded. At the sprint he added
a benchmark using docutils to process 'translation.txt' from pypy's
own documentation and Guido added another using his own 'templess'
Holger worked on getting some code written for bzr that makes nice
graphs out of benchmark to parse the benchmark data produced by PyPy's
Maciej worked on the lib/distributed code that demonstrates the PyPy's
transparent proxies. After a bit of effort, he was able to write a
demo that implemented a remote version of pdb by simply creating a
traceback object that proxied all operations to a remote process.
Michael and Richard spent a day or so on the LLVM backend, which of
late hasn't been so much "maintained" as "held together by
increasingly large amounts of sticky tape". After some refactoring of
the way the backend handled options, they removed a layer of hacks
around the issue of FixedSizedArrays and implemented them properly,
and also added support for the direct_* pointer operations produced by
Michael spent some time using Shark, an OS X profiling application,
and found some OS specific flags and tweaks that improved the
performance of pypy-c on OS X/PPC by around 20%.
As readers of pypy-dev will know by now, there were discussions about
how PyPy is going to continue after the end of the EU funding period.
However, we don't have to summarize them here because we can just link
to Armin's mail:
mwh & Carl Friedrich
More information about the Pypy-dev