Differences between revisions 44 and 45
Revision 44 as of 2006-05-27 15:19:27
Size: 7035
Comment:
Revision 45 as of 2006-05-27 15:50:13
Size: 7146
Comment:
Deletions are marked like this. Additions are marked like this.
Line 154: Line 154:
 * [unfinished] _sre.c matching can be sped up by 3-5% by using free lists for the small frequently allocated objects. It uses very few at the same time - even caching a single SRE_REPEAT object gives a speedup of almost 1%. I [JackDiederich] will try and finish the patch before 2.5beta.  * regexp searching speedup by 1%-2% on linux just by using PyObject_(MALLOC|FREE) for small allocations instead of system malloc.  Unfinished is a free lists implementation for the small frequently allocated objects. _sre.c uses very few objects at the same time - even caching a single SRE_REPEAT object gives a speedup of almost 1%! I [JackDiederich] will try and finish the patch before 2.5beta.

Things we think the Python community will like.

  • Frame optimizations: once a function is called it retains the allocated frame for use in future calls, avoiding allocation and initialization overhead. Frame size has also been slightly reduced (RichardJones).

    This gave a 10% PyStone improvement on RichardJones' test machine, compared to Python 2.4 (from 20242 to 22935).

  • Applied Py_LOCAL and PY_LOCAL_AGGRESSIVE compiler tuning from SRE to ceval.c (currently for Windows only). This results in a 3-10% PyStone speedup on our Intel boxes (FredrikLundh).

  • Made Gzip readline 30-40% faster (BobIppolito)

  • Speed up Unicode operations (AndrewDalke, FredrikLundh). Most notable, repeat is much faster, and most search operations (find, index, count, in) are a LOT faster (25x for the related stringbench tests). Also, rsplit is now as fast as split, and splitlines is nearly as fast as a plain split("\n"). Current stringbench results:

        str(ms) uni(ms) %       comment
        -----------------------------------------------------
        2271.31 3608.32 62.9    TOTAL 2.5a2
        2261.85 1187.84 190.4   TOTAL tuesday
        2247.84  875.13 256.9   TOTAL wednesday
  • (yes, the Unicode string type is now more than twice as fast on this set of tests, and over 4 times faster than when we started. ymmv.)

        str(ms) uni(ms) %       comment
        -----------------------------------------------------
        2271.31 3608.32 62.9    TOTAL 2.5a2
        694.97  883.67  78.6    TOTAL thursday
  • Patch 1335972 was a combination bugfix+speedup for string->int conversion. These are the speedups measured on my Windows box for decimal strings of various lengths. Note that the difference between 9 and 10 is the difference between short and long Python ints on a 32-bit box. (TimPeters)

The patch doesn't actually do anything to speed conversion to long directly; the speedup in those cases is solely due to detecting "unsigned long" overflow more quickly:

        length speedup
        ------ -------
         1       12.4%
         2       15.7%
         3       20.6%
         4       28.1%
         5       33.2%
         6       37.5%
         7       41.9%
         8       46.3%
         9       51.2%
        10       19.5%
        11       19.9%
        12       23.9%
        13       23.7%
        14       23.3%
        15       24.9%
        16       25.3%
        17       28.3%
        18       27.9%
        19       35.7%
  • The struct module has been rewritten to pre-compile struct descriptors (similar to the RE module). This gives a 20% speedup, on average, for the test suite (BobIppolito). Taking advantage of new ability to "compile" a struct pattern (similar to compiling regexps) can be much faster still.

  • Worked on using profile guided optimizations in Visual Studio 8 (KristjanJonsson, RichardMTew). This appears to give on the order of 15% speed improvement in the pybench test suite. A new PCBuild8 directory will be added with automated mechanisms for doing this.

  • Patch 1442927 aimed at speeding long(str, base) operations. It required major fiddling for portability, endcase correctness, and avoiding significant slowdowns on shorter input strings. Now it's up to 6x faster, although it takes a lot of digits to get that, and it's still slower for 1-digit inputs. (TimPeters)

Speedups at various lengths for decimal inputs, comparing 2.4.3 with current trunk:

  len  speedup
 ----  -------
   1     -4.5%
   2      4.6%
   3      8.3%
   4     12.7%
   5     16.9%
   6     28.6%
   7     35.5%
   8     44.3%
   9     46.6%
  10     55.3%
  11     65.7%
  12     77.7%
  13     73.4%
  14     75.3%
  15     85.2%
  16    103.0%
  17     95.1%
  18    112.8%
  19    117.9%
  20    128.3%
  30    174.5%
  40    209.3%
  50    236.3%
  60    254.3%
  70    262.9%
  80    295.8%
  90    297.3%
 100    324.5%
 200    374.6%
 300    403.1%
 400    391.1%
 500    388.7%
 600    440.6%
 700    468.7%
 800    498.0%
 900    507.2%
1000    501.2%
2000    450.2%
3000    463.2%
4000    452.5%
5000    440.6%
6000    439.6%
7000    424.8%
8000    418.1%
9000    417.7%
  • Replaced some Python C-API calls by newer, speedier versions. Most notably, using PyObject_CallFunctionObjArgs instead of PyObject_CallFunction when all arguments are PyObject pointers (the O format code) can shave some 30% off the actual function call (GeorgBrandl).

  • Reworked, debugged and applied a patch to speed up interpreter startup by reducing the number of open() calls (GeorgBrandl).

    • Even for repeated invocation of the interpreter where all files remain in the cache, you get the following figures:

   Time for 1000 interpreter startups
   Python 2.4 trunk  16.470 secs
   Python 2.5.2a2    17.328 secs
   Python 2.5 trunk  15.873 secs
  • Made all built-in exceptions real C new-style types. In 2.5 alpha, they were just objects faking to be classes, which slowed them down by 20% compared to 2.4. After this change, exception handling is around 30% faster than in 2.4 (RichardJones, GeorgBrandl).

  • Added support for min(), max(), sum(), and pow() to Psyco. This produces significant speedups when used in combination with virtualized objects. (JohnBenediktsson)

    Operation             CPython 2.4.3    Psyco 1.5.1   Psyco-NFS
    ------------------    -------------    -----------   -----------
    sum(range(100))       9.856            8.709         0.016
    min(1, max(2, 3))     1.131            0.504         0.017
    pow(3, 2, 2)          0.602            0.140         0.016
  • Added support for int(), float(), and round() to Psyco. These become quite efficient within a set of other psyco operations. (JohnBenediktsson)

    Operation             CPython 2.4.3    Psyco 1.5.1   Psyco-NFS
    ------------------    -------------    -----------   -----------
    int(round(5.375))     0.975            0.461         0.015
  • /!\ NB. These operations now all take 0.016 seconds because they are constant-folded! You need examples with non-constant arguments in the branchmarks. Also, you need to compile Psyco in debugging mode for testing - some of these examples generate fatal assertion errors in the Psyco-NFS branch. (ArminRigo)

  • GeorgBrandl and JackDiederich worked on a rewrite of the Decimal module in C. This will kick-start a Google SummerOfCode project, which will therefore achieve more than originally planned.

  • regexp searching speedup by 1%-2% on linux just by using PyObject_(MALLOC|FREE) for small allocations instead of system malloc. Unfinished is a free lists implementation for the small frequently allocated objects. _sre.c uses very few objects at the same time - even caching a single SRE_REPEAT object gives a speedup of almost 1%! I [JackDiederich] will try and finish the patch before 2.5beta.

NeedForSpeed/Successes (last edited 2008-11-15 14:00:10 by localhost)

Unable to edit the page? See the FrontPage for instructions.