| msg111895 - (view) |
Author: Alexandre Vassalotti (alexandre.vassalotti) *  |
Date: 2010-07-29 06:25 |
| This is a big patch. Please review at http://codereview.appspot.com/1694050/show This patch adds the most interesting optimizations from Unladen Swallow to Python 3's pickle. The core of the patch already been reviewed by Antoine and me (http://codereview.appspot.com/33070/show). One of the last issue remaining the unbounded size of the internal buffer. This shouldn't be a big issue for most uses of pickle, since the size of a pickle is often several times smaller than the object hierarchy that created it. I still hope to fix this in a following patch. The patch also include additional cleanups to the Pdata structure. The changes felt natural to make along with the other changes from Unladen Swallow. IIRC, these changes yield an additional 1-3% speedup. |
|
|
| msg111904 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-07-29 09:41 |
| Le jeudi 29 juillet 2010 à 06:26 +0000, Alexandre Vassalotti a écrit : > New submission from Alexandre Vassalotti <alexandre@peadrop.com>: > > This is a big patch. Please review at > http://codereview.appspot.com/1694050/show > > This patch adds the most interesting optimizations from Unladen > Swallow to Python 3's pickle. > > The core of the patch already been reviewed by Antoine and me > (http://codereview.appspot.com/33070/show). One of the last issue > remaining the unbounded size of the internal buffer. This shouldn't be > a big issue for most uses of pickle, since the size of a pickle is > often several times smaller than the object hierarchy that created it. > I still hope to fix this in a following patch. I still think this should be fixed in this patch, especially since it will also change benchmark numbers. |
|
|
| msg111958 - (view) |
Author: Alexander Belopolsky (belopolsky) * |
Date: 2010-07-29 16:12 |
| For those not familiar with Unladen Swallow, can you describe what the "most interesting optimizations" are? Maybe there is an Unladen Swallow document you can point to. Would any of these optimizations apply to python implementation? |
|
|
| msg111960 - (view) |
Author: STINNER Victor (vstinner) * |
Date: 2010-07-29 16:31 |
| I'm working on #3873 to add a read buffer (fixed size, 4096 bytes) to the unpickler. It's 6 to 8 times faster with the read buffer: but this patch is mainly to avoid the overhead introduced by the new I/O library (in Python2, unpickler was faster because it doesn't need to call Python functions to read some bytes). Is this feature included in this big patch? |
|
|
| msg112091 - (view) |
Author: Skip Montanaro (skip.montanaro) *  |
Date: 2010-07-30 17:38 |
| Look around the issues. I'm pretty sure I worked on the unbounded size issue at one point. Skip |
|
|
| msg112956 - (view) |
Author: Alexandre Vassalotti (alexandre.vassalotti) * |
Date: 2010-08-05 08:13 |
| Ah, that's right Skip. You did fixed it in Unladen Swallow's trunk. I will take a look at your solution. http://code.google.com/p/unladen-swallow/source/diff?spec=svn1167&r=1038&format=side&path=/trunk/Modules/cPickle.c&old_path=/trunk/Modules/cPickle.c&old=988 |
|
|
| msg115718 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-06 17:09 |
| The patch doesn't apply cleanly anymore. Furthermore, I discovered some additional issues: - load, dump, loads and dumps from the _pickle module were never used because they were shadowed by the same functions in pickle.py - once the C functions above are enabled, they produce some bugs I'm working on an updated patch, fixing the aforementioned bugs and adding a buffer size limit. |
|
|
| msg115723 - (view) |
Author: Alexandre Vassalotti (alexandre.vassalotti) * |
Date: 2010-09-06 17:28 |
| Antoine, I fixed these issues in the latest patch posted on Rietveld. Also, Skip added the buffer limit in Unladen Swallow (see ). We just need to merge that. |
|
|
| msg115724 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-06 17:35 |
| Here is a patch. Benchmark numbers: * dumps(): ./python -m timeit -s "import pickle, io; d={(x, 'a'): x for x in range(10000)}" "pickle.dumps(d)" -> before: 100 loops, best of 3: 7.47 msec per loop -> after: 100 loops, best of 3: 2.45 msec per loop * loads(): ./python -m timeit -s "import pickle, io; d={(x, 'a'): x for x in range(10000)}; d=pickle.dumps(d)" "pickle.loads(d)" -> before: 100 loops, best of 3: 12.1 msec per loop -> after: 100 loops, best of 3: 2.62 msec per loop * dump(): ./python -m timeit -s "import pickle, io; d={(x, 'a'): x for x in range(10000)}" "pickle.dump(d, io.BytesIO())" -> before: 100 loops, best of 3: 13.2 msec per loop -> after: 100 loops, best of 3: 2.54 msec per loop * load(): ./python -m timeit -s "import pickle, io; d={(x, 'a'): x for x in range(10000)}; d=pickle.dumps(d)" "pickle.load(io.BytesIO(d))" -> before: 100 loops, best of 3: 12.7 msec per loop -> after: 100 loops, best of 3: 11.6 msec per loop As you can see, load() doesn't really benefit from the buffering improvements. The three methods see quite massive speedups. |
|
|
| msg115725 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-06 17:36 |
| Gosh. My patch is based on an outdated patch :( |
|
|
| msg115726 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-06 18:19 |
| Ok, this patch merges my changes with Alexandre's previous patch. Performance is similar as the previous posted patch. |
|
|
| msg115771 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-07 15:12 |
| For the record, here are the unladen swallow benchmark results against stock py3k: ### pickle ### Min: 1.644222 -> 0.812691: 2.0232x faster Avg: 1.652311 -> 0.814994: 2.0274x faster Significant (t=297.660908) Stddev: 0.00594 -> 0.00207: 2.8732x smaller ### unpickle ### Min: 2.802217 -> 0.751013: 3.7312x faster Avg: 2.807074 -> 0.752646: 3.7296x faster Significant (t=980.311525) Stddev: 0.00446 -> 0.00145: 3.0831x smaller ### pickle_dict ### Min: 0.744259 -> 0.543992: 1.3681x faster Avg: 0.747806 -> 0.545883: 1.3699x faster Significant (t=114.070170) Stddev: 0.00266 -> 0.00293: 1.1014x larger ### pickle_list ### Min: 2.058899 -> 1.212566: 1.6980x faster Avg: 2.066486 -> 1.216711: 1.6984x faster Significant (t=269.964154) Stddev: 0.00534 -> 0.00459: 1.1635x smaller ### unpickle_list ### Min: 1.458531 -> 0.313154: 4.6576x faster Avg: 1.464023 -> 0.319126: 4.5876x faster Significant (t=425.745063) Stddev: 0.00476 -> 0.00367: 1.2976x smaller |
|
|
| msg115957 - (view) |
Author: Antoine Pitrou (pitrou) * |
Date: 2010-09-09 18:33 |
| The patch was finally committed in r84653. Thanks to everyone who participated in this. |
|
|