An instructive lesson in YAGNI ("you aren't going to need it"),
premature optimization, and not making assumptions about Python data
structure implementations.
I need a 1000 x 1000 two-dimensional array of objects. (Since they are
instances of application classes it appears that the array module is
useless; likewise, since I am using Python 3.1, so among other things,
I can't use numpy or its relatives.) The usage pattern is that the
array is first completely filled with objects. Later, objects are
sometimes accessed individually by row and column and often the entire
array is iterated over.
Worried (unnecessarily, as it turns out) by the prospect of 1,000,000
element list I started by constructing a dictionary with the keys 1
through 1000, each of which had as its value another dictionary with
the keys 1 through 1000. Actual values were the values of the second
level dictionary.
Using numbers to fill the array to minimize the effect of creating my
more complex objects, and running Python 3.1.1 on an 8-core Mac Pro
with 8Gb memory, I tried the following
#create and fill the array:
t1 = time.time()
d2 = {}
for j in range(1000):
d2[j] = dict()
for k in range(1000):
d2[j][k] = k
print( round(time.time() - t1, 2))
0.41
# access each element of the array:
t1 = time.time()
for j in range(1000):
for k in range(1000):
elt = d2[j][k]
print( round(time.time() - t1, 2))
0.55
My program was too slow, so I started investigating whether I could
improve on the two-level dictionary, which got used a lot. To get
another baseline I tried a pure 1,000,000-element list, expecting the
times to be be horrendous, but look!
# fill a list using append
t1 = time.time()
lst = []
for n in range(1000000):
lst.append(n)
print( round(time.time() - t1, 2))
0.26
# access every element of a list
t1 = time.time()
for n in range(1000000):
elt = lst[n]
print( round(time.time() - t1, 2))
0.25
What a shock! I could save half the execution time and all my clever
work and awkward double-layer dictionary expressions by just using a
list!
Even better, look what happens using a comprehension to create the
list instead of a loop with list.append:
t1 = time.time()
lst = [n for n in range(1000000)]
print( round(time.time() - t1, 2))
0.11
Half again to create the list.
Iterating over the whole list is easier and faster than iterating over
the double-level dictionary, in particular because it doesn't involve
a two-level loop. But what about individual access given a row and a
column?
t1 = time.time()
for j in range(1000):
for k in range(1000):
elt = lst[j * 1000 + k]
print( round(time.time() - t1, 2))
0.45
This is the same as for the dictionary.
I tried a two-level list and a few other things but still haven't
found anything that works better than a single long list -- just like
2-D arrays are coded in old-style languages, with indices computed as
offsets from the beginning of the linear sequence of all the values.
What's amazing is that creating and accessing 1,000,000-element list
in Python is so efficient. The usual moral obtains: start simple,
analyze problems (functional or performance) as they arise, decide
whether they are worth the cost of change, then change in very limited
ways. And of course abstract and modularize so that, in my case, for
example, none of the program's code would be affected by the change
from a two-level dictionary representation to using a single long list.
I realize there are many directions an analysis like this can follow,
and many factors affecting it, including patterns of use. I just
wanted to demonstrate the basics for a situation that I just
encountered. In particular, if the array was sparse, rather than
completely full, the two-level dictionary implementation would be the
natural representation.
--
http://mail.python.org/mailman/listinfo/python-list
