| 2004-08-27 |
concat vs join - followup |
|
A couple of people have made good points about my last post comparing string concatenation and join.
Marilyn Davis pointed out that in my data, the crossover point where join beats concatenation is always around 500 total characters in the final string. Hans Nowak pointed out that for much longer strings such the lines of a file or parts of a web page, the crossover point comes very quickly.
So here is ConcatTimer version 2 :-) This version dispenses with the fancy graphics and just looks for the crossover point. (It's not too smart about it, either.) It also looks at much larger text chunks - up to 80 characters. Here is the program:
import timeit
reps = 100
unit = ' '
def concatPlus(count):
s=''
for i in range(count):
s += unit
return s
def concatJoin(count):
s=[]
for i in range(count):
s.append(unit)
return ''.join(s)
def timeOne(fn, count):
setup = "from __main__ import " + fn.__name__
stmt = '%s(%d)' % (fn.__name__, count)
t = timeit.Timer(stmt, setup)
secs = min(t.repeat(3, reps))
return secs
def findOne(unitLen):
global unit
unit = ' ' * unitLen
t = 2
while 1:
tPlus = timeOne(concatPlus, t)
tJoin = timeOne(concatJoin, t)
if tPlus > tJoin:
break
t += 1
return t, tPlus, tJoin
for unitLen in range(1,80):
t, tPlus, tJoin = findOne(unitLen)
print '%2d %3d %3d %1.5f %1.5f' % (unitLen, t, t*unitLen, tPlus, tJoin)
And here is an elided list of results. The columns are the length of the pieces, the number of pieces where concat becomes more expensive than join, the total number of characters in the string at the crossover point, and the actual times. (I cut the number of reps down to keep this from taking too long to run.)
1 475 475 0.02733 0.02732
2 263 526 0.01581 0.01581
3 169 507 0.01024 0.01022
4 129 516 0.00782 0.00778
5 100 500 0.00622 0.00604
6 85 510 0.00517 0.00515
7 73 511 0.00447 0.00446
8 63 504 0.00386 0.00385
9 57 513 0.00354 0.00353
10 53 530 0.00333 0.00333
11 47 517 0.00294 0.00292
12 45 540 0.00287 0.00285
13 41 533 0.00262 0.00260
14 38 532 0.00246 0.00244
15 36 540 0.00232 0.00230
16 34 544 0.00222 0.00222
17 31 527 0.00200 0.00199
18 29 522 0.00189 0.00188
19 30 570 0.00199 0.00194
20 28 560 0.00188 0.00186
21 28 588 0.00190 0.00185
22 26 572 0.00177 0.00174
23 25 575 0.00170 0.00168
24 24 576 0.00165 0.00163
25 23 575 0.00158 0.00156
26 22 572 0.00153 0.00151
27 21 567 0.00146 0.00144
28 21 588 0.00146 0.00146
29 21 609 0.00147 0.00144
30 20 600 0.00142 0.00139
31 19 589 0.00134 0.00134
32 20 640 0.00143 0.00139
33 19 627 0.00137 0.00136
34 18 612 0.00130 0.00129
35 18 630 0.00131 0.00130
36 18 648 0.00133 0.00130
37 17 629 0.00126 0.00126
38 17 646 0.00126 0.00124
39 15 585 0.00112 0.00111
43 15 645 0.00113 0.00110
44 14 616 0.00106 0.00105
45 15 675 0.00114 0.00110
46 14 644 0.00106 0.00105
48 14 672 0.00109 0.00105
49 13 637 0.00100 0.00099
58 13 754 0.00104 0.00100
59 12 708 0.00098 0.00096
69 12 828 0.00102 0.00098
70 11 770 0.00093 0.00092
77 11 847 0.00094 0.00091
78 10 780 0.00086 0.00086
79 10 790 0.00087 0.00085
So, for anyone still reading, you can see that Hans is right and Marilyn is close:
- For longer strings and more than a few appends, join is clearly a win
- The total number of characters at the crossover isn't quite constant, but it grows slowly.
Based on this experiment I would say that if the total number of characters is less than 500-1000, concatenation is fine. For anything bigger, use join.
Of course the total amount of time involved in any case is pretty small. Unless you have a lot of characters or you are building a lot of strings, I don't think it really matters too much.
I started this experiment because I have been telling people on the Tutor mailing list to use join, and I wondered how much it really mattered. Does it make enough of a difference to bring it up to beginners? I'm not sure. It's good to teach best practices, but maybe it's a poor use of time to teach this to beginners. I won't be so quick to bring it up next time.
Update: Alan Gauld points out that this is an optimization, and the first rule of optimization is don't until you know you need it. That's a useful way to think about it. Thanks for the reminder!
|
|
posted at 23:39:44
#
comment []
trackback []
|
| |
Which is really faster - concatenation or join? |
|
A couple of times recently I have given out the conventional advice that for concatenating strings, it is better to build a list of the pieces and join it together than to use string concatenation to build the list. The reasoning is that string concatenation requires copying the entire string for each addition, while the list is designed to make concatenation efficient.
Because of all the copying, the time for string concatenation is proportional to the square of the number of additions - it is O(n^2). List append, however, happens in constant time, so building the list takes time proportional to the number of appends - it is O(n).
The trick with this, though, is that there is a proportionality constant here, and for small n, string concatenation may be faster. I decided to find out.
Here is a program that compares the time for string concatenation using the two methods. It varies both the number of append operations and the length of the appended strings. It prints the results in a series of tables, and it uses VPython to graph the results:
import timeit
from visual.graph import *
reps = 1000
unit = ' '
def concatPlus(count):
s=''
for i in range(count):
s += unit
return s
def concatJoin(count):
s=[]
for i in range(count):
s.append(unit)
return ''.join(s)
def timeOne(fn, count):
setup = "from __main__ import " + fn.__name__
stmt = '%s(%d)' % (fn.__name__, count)
t = timeit.Timer(stmt, setup)
secs = min(t.repeat(3, reps))
return secs
def graphOne(unitLen):
global unit
unit = ' ' * unitLen
title = 'Unit length is %d' % len(unit)
funct1 = gcurve(color=color.cyan)
funct2 = gdots(color=color.yellow)
print
print title
print ' tPlus tJoin'
for t in range(10, 100, 10) + range(100, 600, 50):
tPlus = timeOne(concatPlus, t)
tJoin = timeOne(concatJoin, t)
print '%5d %2.3f %2.3f' % (t, tPlus, tJoin)
funct1.plot( pos=(t, tPlus) )
funct2.plot( pos=(t, tJoin) )
graph = gdisplay(title='Append speed', xtitle='count', ytitle='time')
for unitLen in [1,2,3,4,5]:
graphOne(unitLen)
Here is the graph - the yellow dots are for concatJoin, the blue curves are concatPlus:
A couple of things stand out from this:
- For every string length, concatPlus is faster when the number of appends is relatively small - up to 80 appends in my tests
- For larger numbers of appends, not only does concatPlus show O(n^2) behavior, it gets worse as the size of the appended strings grows. concatJoin is O(n) and it doesn't really matter how long the appended string is. In fact, I think concatPlus is O(m*n) where m is the total length of the final string and n is the number of appends.
Based on these results, I think I will stop spreading the conventional wisdom. I think most uses of string concatenation are for small strings with a small number of concatenations. String join only pays off when there are a lot of appends.
Here is the raw data from the program:
D:\Personal\Tutor>python concattimer.py
Visual-2003-10-05
Unit length is 1
tPlus tJoin
10 0.005 0.008
20 0.008 0.014
30 0.012 0.020
40 0.015 0.025
50 0.018 0.032
60 0.022 0.037
70 0.026 0.044
80 0.029 0.049
90 0.033 0.057
100 0.038 0.062
150 0.059 0.092
200 0.082 0.122
250 0.109 0.149
300 0.145 0.178
350 0.184 0.208
400 0.222 0.237
450 0.262 0.264
500 0.307 0.292
550 0.349 0.325
Unit length is 2
tPlus tJoin
10 0.005 0.008
20 0.008 0.014
30 0.012 0.019
40 0.015 0.026
50 0.019 0.033
60 0.023 0.039
70 0.027 0.044
80 0.033 0.051
90 0.038 0.057
100 0.042 0.062
150 0.075 0.095
200 0.114 0.125
250 0.155 0.154
300 0.200 0.185
350 0.247 0.215
400 0.295 0.243
450 0.349 0.272
500 0.404 0.305
550 0.463 0.332
Unit length is 3
tPlus tJoin
10 0.005 0.008
20 0.008 0.014
30 0.012 0.019
40 0.016 0.026
50 0.020 0.033
60 0.025 0.039
70 0.031 0.045
80 0.036 0.051
90 0.043 0.057
100 0.050 0.064
150 0.090 0.095
200 0.134 0.127
250 0.184 0.159
300 0.236 0.187
350 0.293 0.217
400 0.360 0.247
450 0.427 0.278
500 0.499 0.306
550 0.576 0.335
Unit length is 4
tPlus tJoin
10 0.005 0.008
20 0.008 0.014
30 0.012 0.019
40 0.017 0.025
50 0.021 0.033
60 0.026 0.039
70 0.035 0.045
80 0.042 0.051
90 0.050 0.058
100 0.057 0.063
150 0.101 0.096
200 0.153 0.129
250 0.207 0.161
300 0.271 0.192
350 0.341 0.222
400 0.416 0.249
450 0.501 0.280
500 0.580 0.310
550 0.682 0.338
Unit length is 5
tPlus tJoin
10 0.005 0.008
20 0.008 0.014
30 0.013 0.019
40 0.017 0.025
50 0.022 0.034
60 0.029 0.040
70 0.040 0.046
80 0.047 0.052
90 0.055 0.058
100 0.063 0.064
150 0.114 0.097
200 0.169 0.130
250 0.232 0.163
300 0.306 0.195
350 0.385 0.223
400 0.470 0.252
450 0.567 0.283
500 0.668 0.313
550 0.779 0.344
|
|
posted at 19:49:20
#
comment []
trackback []
|
|
|
| August 2004 |
|---|
| Mo | Tu | We | Th | Fr | Sa | Su |
|---|
| | | | | | | 1 | | 2 | 3 | 4 | 5 | 6 | 7 | 8 | | 9 | 10 | 11 | 12 | 13 | 14 | 15 | | 16 | 17 | 18 | 19 | 20 | 21 | 22 | | 23 | 24 | 25 | 26 | 27 | 28 | 29 | | 30 | 31 | | | | | | Jun
2004 | | Sep
2004 |
|---|
Comments about life, the universe and Python, from the imagination of Kent S Johnson.
BlogRoll
|
