constant in python?

Bengt Richter bokr at accessone.com
Sat Aug 18 21:22:32 CEST 2001


On Sat, 18 Aug 2001 18:59:30 +0200, "Alex Martelli" <aleaxit at yahoo.com> wrote:

>"Michael Ströder" <michael at stroeder.com> wrote in message
>news:3B7E5355.675D569B at stroeder.com...
>> Alex Martelli wrote:
>> >
>> > > > http://aspn.activestate.com/ASPN/Cookbook/Python/Recipe/65207
>> >
>> > If something can be done with a few lines of code in the
>> > language as it is, it requires very strong motivation for that
>> > something to become a built-in language mechanism:
>>
>> I would guess that the recipe above causes some performance penalty.
>
>Why guess, when you can measure?
>
>Say const.py is like in the recipe, normal.py is an empty
>normal module, and test.py contains:
>
>import time, const, normal
>
>start = time.clock()
>const.magic = 23
>sum = 0L
>for i in range(100000):
>    sum += const.magic
>stend = time.clock()
>
>print "const : %5.2f"%(stend-start)
>
>start = time.clock()
>normal.magic = 23
>sum = 0L
>for i in range(100000):
>    sum += normal.magic
>stend = time.clock()
>
>print "normal: %5.2f"%(stend-start)
>
>Here are three runs on my box:
>
>D:\ian\good>python test.py
>const :  1.44
>normal:  1.60
>
>D:\ian\good>python test.py
>const :  1.63
>normal:  1.53
>
>D:\ian\good>python test.py
>const :  1.41
>normal:  1.53
>
>See?  No statistically significant difference penalizes the
>const module vs a normal module -- it so happens the
>loop using the const module was faster two times out
>of three, but this is indeed happenstance.
>
>Moral: never guess at performance -- it often baffles the
>intuition even of the most accomplished expert.  MEASURE.
>
I agree, but the happenstances can be separated out to a fair degree,
even if you're on line and background stuff may be happening:

(Not very elegant cutting and pasting, but here's what I like to do :)
_______________________________________________

import time, const, normal

const.magic = 23
normal.magic = 23
sumc = 0L
sumn = 0L
cdir={}
ndir={}
for i in range(100000):
    start = time.clock()
    sumc += const.magic
    stend = time.clock()
    sumn += normal.magic
    stend2 = time.clock()
    dt = int((stend-start)*1.0e9)	#nanoseconds
    if cdir.has_key(dt):
        cdir[dt] += 1
    else:
	cdir[dt] = 1

    dt = int((stend2-stend)*1.0e9)	#nanoseconds
    if ndir.has_key(dt):
        ndir[dt] += 1
    else:
	ndir[dt] = 1

print '-- const case --'
vk = [(v,k) for k,v in cdir.items()]
vk.sort()
vk.reverse()
for v,k in vk[:10]:
    print "%7d %10d" % ( v, k )

print '-- normal case --'
vk = [(v,k) for k,v in ndir.items()]
vk.sort()
vk.reverse()
for v,k in vk[:10]:
    print "%7d %10d" % ( v, k )

__________________
Result: (NT4sp3,300mhzP2,320MB ram, Python 2.1)

-- const case --
  54910      31847
  21574      31009
  20695      32685
   1883      33523
    327      30171
    119      34361
     72      35199
     54      38552
     49      39390
     28      36038
-- normal case --
  55261      30171
  26861      31009
  14372      29333
   2785      31847
    305      32685
     51      37714
     41      28495
     31      38552
     27      36876
     19      39390

But note that you don't have to run 100k trials
to get useful results, since you see the outliers
that might otherwise get averaged in. Here's for 1k:

-- const case --
    549      31009
    293      30171
    135      31847
     11      32685
      6      29333
      1    1063542
      1     364571
      1      46095
      1      35199
      1      34361
-- normal case --
    646      29333
    205      28495
    128      30171
     13      31009
      1     402285
      1      99733
      1      93028
      1      68723
      1      63695
      1      54476




More information about the Python-list mailing list