profile, cProfile, and pstats – Performance analysis of Python programs.¶
| Purpose: | Performance analysis of Python programs. |
|---|---|
| Available In: | 1.4 and later, these examples are for Python 2.5 |
The profile and cProfile modules provide APIs for collecting and analyzing statistics about how Python source consumes processor resources.
run()¶
The most basic starting point in the profile module is run(). It takes a string statement as argument, and creates a report of the time spent executing different lines of code while running the statement.
import profile
def fib(n):
# from http://en.literateprograms.org/Fibonacci_numbers_(Python)
if n == 0:
return 0
elif n == 1:
return 1
else:
return fib(n-1) + fib(n-2)
def fib_seq(n):
seq = [ ]
if n > 0:
seq.extend(fib_seq(n-1))
seq.append(fib(n))
return seq
print 'RAW'
print '=' * 80
profile.run('print fib_seq(20); print')
This recursive version of a fibonacci sequence calculator [1] is especially useful for demonstrating the profile because we can improve the performance so much. The standard report format shows a summary and then details for each function executed.
$ python profile_fibonacci_raw.py
RAW
================================================================================
[0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
57356 function calls (66 primitive calls) in 0.746 CPU seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
21 0.000 0.000 0.000 0.000 :0(append)
20 0.000 0.000 0.000 0.000 :0(extend)
1 0.001 0.001 0.001 0.001 :0(setprofile)
1 0.000 0.000 0.744 0.744 <string>:1(<module>)
1 0.000 0.000 0.746 0.746 profile:0(print fib_seq(20); print)
0 0.000 0.000 profile:0(profiler)
57291/21 0.743 0.000 0.743 0.035 profile_fibonacci_raw.py:13(fib)
21/1 0.001 0.000 0.744 0.744 profile_fibonacci_raw.py:22(fib_seq)
As you can see, it takes 57356 separate function calls and 3/4 of a second to run. Since there are only 66 primitive calls, we know that the vast majority of those 57k calls were recursive. The details about where time was spent are broken out by function in the listing showing the number of calls, total time spent in the function, time per call (tottime/ncalls), cumulative time spent in a function, and the ratio of cumulative time to primitive calls.
Not surprisingly, most of the time here is spent calling fib() repeatedly. We can add a memoize decorator [2] to reduce the number of recursive calls and have a big impact on the performance of this function.
import profile
class memoize:
# from http://avinashv.net/2008/04/python-decorators-syntactic-sugar/
def __init__(self, function):
self.function = function
self.memoized = {}
def __call__(self, *args):
try:
return self.memoized[args]
except KeyError:
self.memoized[args] = self.function(*args)
return self.memoized[args]
@memoize
def fib(n):
# from http://en.literateprograms.org/Fibonacci_numbers_(Python)
if n == 0:
return 0
elif n == 1:
return 1
else:
return fib(n-1) + fib(n-2)
def fib_seq(n):
seq = [ ]
if n > 0:
seq.extend(fib_seq(n-1))
seq.append(fib(n))
return seq
if __name__ == '__main__':
print 'MEMOIZED'
print '=' * 80
profile.run('print fib_seq(20); print')
By remembering the Fibonacci value at each level we can avoid most of the recursion and drop down to 145 calls that only take 0.003 seconds. Also notice that the ncalls count for fib() shows that it never recurses.
$ python profile_fibonacci_memoized.py
MEMOIZED
================================================================================
[0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
145 function calls (87 primitive calls) in 0.003 CPU seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
21 0.000 0.000 0.000 0.000 :0(append)
20 0.000 0.000 0.000 0.000 :0(extend)
1 0.001 0.001 0.001 0.001 :0(setprofile)
1 0.000 0.000 0.002 0.002 <string>:1(<module>)
1 0.000 0.000 0.003 0.003 profile:0(print fib_seq(20); print)
0 0.000 0.000 profile:0(profiler)
59/21 0.001 0.000 0.001 0.000 profile_fibonacci_memoized.py:19(__call__)
21 0.000 0.000 0.001 0.000 profile_fibonacci_memoized.py:26(fib)
21/1 0.001 0.000 0.002 0.002 profile_fibonacci_memoized.py:36(fib_seq)
runctx()¶
Sometimes, instead of constructing a complex expression for run(), it is easier to build a simple expression and pass it parameters through a context, using runctx().
import profile
from profile_fibonacci_memoized import fib, fib_seq
if __name__ == '__main__':
profile.runctx('print fib_seq(n); print', globals(), {'n':20})
In this example, the value of “n” is passed through the local variable context instead of being embedded directly in the statement passed to runctx().
$ python profile_runctx.py
[0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
145 function calls (87 primitive calls) in 0.003 CPU seconds
Ordered by: standard name
ncalls tottime percall cumtime percall filename:lineno(function)
21 0.000 0.000 0.000 0.000 :0(append)
20 0.000 0.000 0.000 0.000 :0(extend)
1 0.001 0.001 0.001 0.001 :0(setprofile)
1 0.000 0.000 0.002 0.002 <string>:1(<module>)
1 0.000 0.000 0.003 0.003 profile:0(print fib_seq(n); print)
0 0.000 0.000 profile:0(profiler)
59/21 0.001 0.000 0.001 0.000 profile_fibonacci_memoized.py:19(__call__)
21 0.000 0.000 0.001 0.000 profile_fibonacci_memoized.py:26(fib)
21/1 0.001 0.000 0.002 0.002 profile_fibonacci_memoized.py:36(fib_seq)
pstats: Saving and Working With Statistics¶
The standard report created by the profile functions is not very flexible. If it doesn’t meet your needs, you can produce your own reports by saving the raw profiling data from run() and runctx() and processing it separately with the Stats class from pstats.
For example, to run several iterations of the same test and combine the results, you could do something like this:
import profile
import pstats
from profile_fibonacci_memoized import fib, fib_seq
# Create 5 set of stats
filenames = []
for i in range(5):
filename = 'profile_stats_%d.stats' % i
profile.run('print %d, fib_seq(20)' % i, filename)
# Read all 5 stats files into a single object
stats = pstats.Stats('profile_stats_0.stats')
for i in range(1, 5):
stats.add('profile_stats_%d.stats' % i)
# Clean up filenames for the report
stats.strip_dirs()
# Sort the statistics by the cumulative time spent in the function
stats.sort_stats('cumulative')
stats.print_stats()
The output report is sorted in descending order of cumulative time spent in the function and the directory names are removed from the printed filenames to conserve horizontal space.
$ python profile_stats.py
0 [0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
1 [0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
2 [0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
3 [0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
4 [0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, 610, 987, 1597, 2584, 4181, 6765]
Sun Aug 31 11:29:36 2008 profile_stats_0.stats
Sun Aug 31 11:29:36 2008 profile_stats_1.stats
Sun Aug 31 11:29:36 2008 profile_stats_2.stats
Sun Aug 31 11:29:36 2008 profile_stats_3.stats
Sun Aug 31 11:29:36 2008 profile_stats_4.stats
489 function calls (351 primitive calls) in 0.008 CPU seconds
Ordered by: cumulative time
ncalls tottime percall cumtime percall filename:lineno(function)
5 0.000 0.000 0.007 0.001 <string>:1(<module>)
105/5 0.004 0.000 0.007 0.001 profile_fibonacci_memoized.py:36(fib_seq)
1 0.000 0.000 0.003 0.003 profile:0(print 0, fib_seq(20))
143/105 0.001 0.000 0.002 0.000 profile_fibonacci_memoized.py:19(__call__)
1 0.000 0.000 0.001 0.001 profile:0(print 4, fib_seq(20))
1 0.000 0.000 0.001 0.001 profile:0(print 1, fib_seq(20))
1 0.000 0.000 0.001 0.001 profile:0(print 2, fib_seq(20))
1 0.000 0.000 0.001 0.001 profile:0(print 3, fib_seq(20))
21 0.000 0.000 0.001 0.000 profile_fibonacci_memoized.py:26(fib)
100 0.001 0.000 0.001 0.000 :0(extend)
105 0.001 0.000 0.001 0.000 :0(append)
5 0.001 0.000 0.001 0.000 :0(setprofile)
0 0.000 0.000 profile:0(profiler)
Limiting Report Contents¶
Since we are studying the performance of fib() and fib_seq(), we can also restrict the output report to only include those functions using a regular expression to match the filename:lineno(function) values we want.
import profile
import pstats
from profile_fibonacci_memoized import fib, fib_seq
# Read all 5 stats files into a single object
stats = pstats.Stats('profile_stats_0.stats')
for i in range(1, 5):
stats.add('profile_stats_%d.stats' % i)
stats.strip_dirs()
stats.sort_stats('cumulative')
# limit output to lines with "(fib" in them
stats.print_stats('\(fib')
The regular expression includes a literal left paren (() to match against the function name portion of the location value.
$ python profile_stats_restricted.py
Sun Aug 31 11:29:36 2008 profile_stats_0.stats
Sun Aug 31 11:29:36 2008 profile_stats_1.stats
Sun Aug 31 11:29:36 2008 profile_stats_2.stats
Sun Aug 31 11:29:36 2008 profile_stats_3.stats
Sun Aug 31 11:29:36 2008 profile_stats_4.stats
489 function calls (351 primitive calls) in 0.008 CPU seconds
Ordered by: cumulative time
List reduced from 13 to 2 due to restriction <'\\(fib'>
ncalls tottime percall cumtime percall filename:lineno(function)
105/5 0.004 0.000 0.007 0.001 profile_fibonacci_memoized.py:36(fib_seq)
21 0.000 0.000 0.001 0.000 profile_fibonacci_memoized.py:26(fib)
Caller / Callee Graphs¶
Stats also includes methods for printing the callers and callees of functions.
import profile
import pstats
from profile_fibonacci_memoized import fib, fib_seq
# Read all 5 stats files into a single object
stats = pstats.Stats('profile_stats_0.stats')
for i in range(1, 5):
stats.add('profile_stats_%d.stats' % i)
stats.strip_dirs()
stats.sort_stats('cumulative')
print 'INCOMING CALLERS:'
stats.print_callers('\(fib')
print 'OUTGOING CALLEES:'
stats.print_callees('\(fib')
The arguments to print_callers() and print_callees() work the same as the restriction arguments to print_stats(). The output shows the caller, callee, and cumulative time.
$ python profile_stats_callers.py
INCOMING CALLERS:
Ordered by: cumulative time
List reduced from 13 to 2 due to restriction <'\\(fib'>
Function was called by...
profile_fibonacci_memoized.py:36(fib_seq) <- <string>:1(<module>)(5) 0.007
profile_fibonacci_memoized.py:36(fib_seq)(100) 0.007
profile_fibonacci_memoized.py:26(fib) <- profile_fibonacci_memoized.py:19(__call__)(21) 0.002
OUTGOING CALLEES:
Ordered by: cumulative time
List reduced from 13 to 2 due to restriction <'\\(fib'>
Function called...
profile_fibonacci_memoized.py:36(fib_seq) -> :0(append)(105) 0.001
:0(extend)(100) 0.001
profile_fibonacci_memoized.py:19(__call__)(105) 0.002
profile_fibonacci_memoized.py:36(fib_seq)(100) 0.007
profile_fibonacci_memoized.py:26(fib) -> profile_fibonacci_memoized.py:19(__call__)(38) 0.002
See also
- profile and cProfile
- Standard library documentation for this module.
- pstats
- Standard library documentation for pstats.
- Gprof2Dot
- Visualization tool for profile output data.
| [1] | Fibonacci numbers (Python) - LiteratePrograms |
| [2] | Python Decorators: Syntactic Sugar | avinash.vora |
