The attached patch adds another version of cPickle.c's batch_dict(),
batch_dict_exact(), which is specialized for "type(x) is dict". This
provides a nice performance boost when pickling objects that use
dictionaries:

Pickle:
Min: 2.216 -> 1.858: 19.24% faster
Avg: 2.238 -> 1.889: 18.50% faster
Significant (t=106.874099, a=0.95)

Benchmark is at
http://code.google.com/p/unladen-swallow/source/browse/tests/performance/macro_pickle.py
(driver is ../perf.py; perf.py was run with "--rigorous -b pickle").

This patch passes all the tests added in bpo-5665. I would recommend
reviewing that patch first. I'll port to py3k once this is reviewed for
trunk.

Without taking a very detailed look, the patch looks good.
Are there already tests for pickling of dict subclasses? Otherwise, they
should be added.

By the way, could the same approach be applied to lists and sets as well?

On Thu, Apr 2, 2009 at 12:20 PM, Antoine Pitrou <report@bugs.python.org> wrote:

Antoine Pitrou <pitrou@free.fr> added the comment:

By the way, could the same approach be applied to lists and sets as well?

Certainly; see http://bugs.python.org/issue5671 for the list version.
It doesn't make as big an impact on the benchmark, though.

Certainly; see http://bugs.python.org/issue5671 for the list version.
It doesn't make as big an impact on the benchmark, though.

How about splitting the benchmark in parts:

• (un)pickling lists
• (un)pickling dicts
• (un)pickling sets
  (etc.)

Antoine: pickletester.py:test_newobj_generic() appears to test dict
subclasses, though in a roundabout-ish way. I don't know of any tests
for dict subclasses in the C level sense (ie, PyDict_Check() vs
PyDict_CheckExact()). I can add more explicit tests for Python-level
dict subclasses, if you want.

The patch produces different output for an empty dict: a sequence "MARK
SETITEMS" is written, which is useless and wastes 2 bytes.

Antoine: pickletester.py:test_newobj_generic() appears to test dict
subclasses, though in a roundabout-ish way. I don't know of any tests
for dict subclasses in the C level sense (ie, PyDict_Check() vs
PyDict_CheckExact()). I can add more explicit tests for Python-level
dict subclasses, if you want.

Well, Python-level dict subclasses are also C-level subclasses (in the
PyDict_Check() sense), or am I mistaken?

I ported the patch to py3k. In addition, I added a special-case when the
dict contains only one item; you probably want this special-case in the
trunk version as well.

Oops, I forgot to add the comment on top of batch_dict_exact in the
patch. Here is a better patch.

Oops again, I just remarked that the comment for batch_dict_exact refers
to batch_dict as being above, but I copied batch_dict_exact before
batch_dict. Here's a good patch (hopefully) that puts batch_dict_exact
at the right place.

Silly me, I had changed the PyDict_Size call in outer loop for Py_SIZE
and this is of course totally wrong. Here's a good patch (I am pretty
sure now! ;-) I ran the whole test suite and I saw no failures.

Collin, you can go ahead and commit both patches. Nice work!

Sigh... silly me again. There is some other junk in my last patch.

FYI, I just added a pickle_dict microbenchmark to perf.py. Using this
new microbenchmark, I see these results (perf.py -r -b pickle_dict):

pickle_dict:
Min: 2.092 -> 1.341: 56.04% faster
Avg: 2.126 -> 1.360: 56.37% faster
Significant (t=216.895643, a=0.95)

I still need to address the comment about pickling empty dicts.

Amaury, I can't reproduce the issue you're seeing with empty dicts.
Here's what I'm doing:

dhcp-172-19-19-199:trunk collinwinter$ ./python.exe 
Python 2.7a0 (trunk:71100M, Apr  3 2009, 14:40:49) 
[GCC 4.0.1 (Apple Inc. build 5490)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import cPickle, pickletools
>>> data = cPickle.dumps({}, protocol=2)
>>> pickletools.dis(data)
    0: \x80 PROTO      2
    2: }    EMPTY_DICT
    3: .    STOP
highest protocol among opcodes = 2
>>> data
'\x80\x02}.'
>>>

What are you doing to produce the MARK SETITEMS sequence?

Sorry, I was wrong. I think I noticed that the case size==1 was handled
differently, and incorrectly inferred the same for size==0.
(btw, the patch for trunk was not updated)

Can this patch be used or ported to 2.5.x?

Sorry, it won't even be integrated in 2.6 actually. It's a new feature,
not a bug fix.

Fixed the len(d) == 1 size regression. Final performance of the patch
relative to trunk:

Using Unladen Swallow's perf.py -b pickle,pickle_dict on trunk:
pickle:
Min: 2.238 -> 1.895: 18.08% faster
Avg: 2.241 -> 1.898: 18.04% faster
Significant (t=282.066701, a=0.95)

pickle_dict:
Min: 2.163 -> 1.375: 57.36% faster
Avg: 2.168 -> 1.376: 57.50% faster
Significant (t=527.668441, a=0.95)

Performance for py3k:
pickle:
Min: 2.849 -> 2.790: 2.10% faster
Avg: 2.854 -> 2.796: 2.09% faster
Significant (t=27.624303, a=0.95)

pickle_dict:
Min: 2.121 -> 1.512: 40.27% faster
Avg: 2.128 -> 1.519: 40.13% faster
Significant (t=283.406572, a=0.95)

regrtest.py -uall test_xpickle passes all backwards-compatibility tests
for trunk, and all other tests run by regrtest.py on Linux pass.

Committed as r72909 (trunk), r72910 (py3k).

Thanks!

Committed as r72909 (trunk), r72910 (py3k).

----------
resolution: accepted -> fixed
status: open -> closed