Author vstinner
Recipients christian.heimes, gregory.p.smith, mark.dickinson, vstinner
Date 2008-11-06.13:24:24
SpamBayes Score 8.50986e-14
Marked as misclassified No
Message-id <200811061423.54059.victor.stinner@haypocalc.com>
In-reply-to <200811061346.45179.victor.stinner@haypocalc.com>
Content
And now the stat of Python patched with 30bit_longdigit3.patch.

min/avg/max are now the number of bits which gives better 
informations. "bigger" is the number of arguments which are bigger than 1 
digit (not in range [-2^30; 2^30]).

make
====

_FromLong:   169734 calls, min=( 0,  ), avg=(11.6,     ), max=(  32,     )
  \--> bigger=31086
long_bool:    48772 calls, min=( 0,  ), avg=( 0.3,     ), max=(  24,     )
long_add:     39685 calls, min=( 0, 0), avg=( 6.5,  3.5), max=(  19,   32)
  \--> bigger=1
long_compare: 39445 calls, min=( 0, 0), avg=( 9.3,  8.4), max=(  31,   33)
  \--> bigger=10438
_AsLong:      33726 calls, min=( 0,  ), avg=( 4.9,     ), max=(1321,     )
  \--> bigger=10
long_sub:     13285 calls, min=( 0, 0), avg=( 7.6,  5.6), max=(  13,   13)
long_bitwise:  4690 calls, min=( 0, 0), avg=( 1.7,  1.9), max=(  16,   16)
long_hash:     1097 calls, min=( 0,  ), avg=( 8.1,     ), max=(  33,     )
  \--> bigger=4
long_mul:       236 calls, min=( 0, 0), avg=( 1.3,  5.4), max=(  17,   17)
long_invert:    204 calls, min=( 0,  ), avg=( 2.4,     ), max=(   3,     )
long_neg:        35 calls, min=( 1,  ), avg=( 4.3,     ), max=(   7,     )
long_format:      3 calls, min=( 0,  ), avg=( 2.0,     ), max=(   4,     )
long_mod:         3 calls, min=( 1, 2), avg=( 1.7,  2.0), max=(   2,    2)
long_pow:         1 calls, min=( 2, 6), avg=( 2.0,  6.0), max=(   2,    6)

Notes about make:
 - PyLong_FromLong(), long_compare(), PyLong_AsLong() and long_hash() 
   gets integers not in [-2^30; 2^30] which means that all other functions
   are only called with arguments of 1 digit!
 - PyLong_FromLong() gets ~30.000 (18%) integers of 32 bits
 - global average integer size is between 0.3 and 11.6 (~6.0 bits?)
 - There are 41.500 (12%) big integers on ~350.000 integers

pystone
=======

_FromLong:   1504983 calls, min=( 0,  ), avg=( 5.1,     ), max=(  31,     )
  \--> bigger=14
long_add:     902487 calls, min=( 0, 0), avg=( 3.9,  2.4), max=(  17,   17)
long_compare: 651165 calls, min=( 0, 0), avg=( 1.7,  1.4), max=(  31,   31)
  \--> bigger=27
_AsLong:      252477 calls, min=( 0,  ), avg=( 4.6,     ), max=(  16,     )
long_sub:     250032 calls, min=( 1, 0), avg=( 4.0,  1.6), max=(   7,    7)
long_bool:    102655 calls, min=( 0,  ), avg=( 0.5,     ), max=(   7,     )
long_mul:     100015 calls, min=( 0, 0), avg=( 2.5,  2.0), max=(   4,   16)
long_truediv:  50000 calls, min=( 4, 2), avg=( 4.0,  2.0), max=(   4,    2)
long_hash:       382 calls, min=( 0,  ), avg=( 8.1,     ), max=(  28,     )
long_bitwise:    117 calls, min=( 0, 0), avg=( 6.7,  6.6), max=(  15,   16)
long_format:       1 calls, min=(16,  ), avg=(16.0,     ), max=(  16,     )

Notes about pystone:
 - very very few numbers are bigger than one digit: 41 / ~4.000.000
 - global average integer size is between 0.5 and 6.7 (~3.0 bits?)
 - the biggest number has only 31 bits (see long_compare)

Short summary:
 - pystone doesn't use big integer (1 big integer for 100.000 integers)
   => don't use pystone!
 - the average integer size is around 3 or 6 bits, which means that most
   integers can be stored in 8 bits (-255..255)
   => we need to focus on the very small numbers
   => base 2^30 doesn't help for common Python code, it only helps programs
      using really big numbers (128 bits or more?)
History
Date User Action Args
2008-11-06 13:24:25vstinnersetrecipients: + vstinner, gregory.p.smith, mark.dickinson, christian.heimes
2008-11-06 13:24:24vstinnerlinkissue4258 messages
2008-11-06 13:24:24vstinnercreate