Line-granularity, thread-aware deterministic and statistic pure-python profiler
Inspired from Robert Kern's line_profiler .
Usage
As a command:
$ pprofile some_python_executable arg1 ...
Once some_python_executable returns, prints annotated code of each file involved in the execution.
As a command, ignoring any files from default sys.path (ie, python modules themselves), for shorter output:
$ pprofile --exclude-syspath some_python_executable arg1 ...
Executing a module, like python -m. --exclude-syspath is not
recommended in this mode, as it will likely hide what you intend to profile.
Also, explicitly ending pprofile arguments with -- will prevent accidentally
stealing command's arguments:
$ pprofile -m some_python_module -- arg1 ...
As a module:
import pprofile
def someHotSpotCallable():
# Deterministic profiler
prof = pprofile.Profile()
with prof():
# Code to profile
prof.print_stats()
def someOtherHotSpotCallable():
# Statistic profiler
prof = pprofile.StatisticalProfile()
with prof(
period=0.001, # Sample every 1ms
single=True, # Only sample current thread
):
# Code to profile
prof.print_stats()For advanced usage, see pprofile --help and pydoc pprofile.
Profiling overhead
pprofile default mode (Deterministic profiling) has a large overhead. Part of the reason being that it is written to be as portable as possible (so no C extension). This large overhead can be an issue, which can be avoided by using Statistic profiling at the cost of some result readability decrease.
Rule of thumb:
| Code to profile runs for... | Deterministic profiling | Statistic profiling |
|---|---|---|
| a few seconds | Yes | No [1] |
| a few minutes | Maybe | Yes |
| more (ex: daemon) | No | Yes [2] |
Once you identified the hot spot and you decide you need finer-grained profiling to understand what needs fixing, you should try to make to-profile code run for shorter time so you can reasonably use deterministic profiling: use a smaller data set triggering the same code path, modify the code to only enable profiling on small pieces of code...
| [1] | Statistic profiling will not have time to collect enough samples to produce usable output. |
| [2] | You may want to consider triggering pprofile from a signal handler or other IPC mechanism to profile a shorter subset. See zpprofile.py for how it can be used to profile code inside a running (zope) service (in which case the IPC mechanism is just Zope normal URL handling). |
Output
Supported output formats.
Callgrind
The most useful output mode of pprofile is Callgrind Profile Format, allows browsing profiling results with kcachegrind (or qcachegrind on Windows).
$ pprofile --format callgrind --out cachegrind.out.threads demo/threads.py
Callgrind format is implicitly enabled if --out basename starts with
cachegrind.out., so above command can be simplified as:
$ pprofile --out cachegrind.out.threads demo/threads.py
If you are analyzing callgrind traces on a different machine, you may want to
use the --zipfile option to generate a zip file containing all files:
$ pprofile --out cachegrind.out.threads --zipfile threads_source.zip demo/threads.py
Generated files will use relative paths, so you can extract generated archive in the same path as profiling result, and kcachegrind will load them - and not your system-wide files, which may differ.
Annotated code
Human-readable output, but can become difficult to use with large programs.
$ pprofile demo/threads.py
Profiling modes
Deterministic profiling
In deterministic profiling mode, pprofile gets notified of each executed line. This mode generates very detailed reports, but at the cost of a large overhead. Also, profiling hooks being per-thread, either profiling must be enable before spawning threads (if you want to profile more than just the current thread), or profiled application must provide ways of enabling profiling afterwards - which is not very convenient.
$ pprofile --threads 0 demo/threads.py
Command line: ['demo/threads.py']
Total duration: 1.00573s
File: demo/threads.py
File duration: 1.00168s (99.60%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
1| 2| 3.21865e-05| 1.60933e-05| 0.00%|import threading
2| 1| 5.96046e-06| 5.96046e-06| 0.00%|import time
3| 0| 0| 0| 0.00%|
4| 2| 1.5974e-05| 7.98702e-06| 0.00%|def func():
5| 1| 1.00111| 1.00111| 99.54%| time.sleep(1)
6| 0| 0| 0| 0.00%|
7| 2| 2.00272e-05| 1.00136e-05| 0.00%|def func2():
8| 1| 1.69277e-05| 1.69277e-05| 0.00%| pass
9| 0| 0| 0| 0.00%|
10| 1| 1.81198e-05| 1.81198e-05| 0.00%|t1 = threading.Thread(target=func)
(call)| 1| 0.000610828| 0.000610828| 0.06%|# /usr/lib/python2.7/threading.py:436 __init__
11| 1| 1.52588e-05| 1.52588e-05| 0.00%|t2 = threading.Thread(target=func)
(call)| 1| 0.000438929| 0.000438929| 0.04%|# /usr/lib/python2.7/threading.py:436 __init__
12| 1| 4.79221e-05| 4.79221e-05| 0.00%|t1.start()
(call)| 1| 0.000843048| 0.000843048| 0.08%|# /usr/lib/python2.7/threading.py:485 start
13| 1| 6.48499e-05| 6.48499e-05| 0.01%|t2.start()
(call)| 1| 0.00115609| 0.00115609| 0.11%|# /usr/lib/python2.7/threading.py:485 start
14| 1| 0.000205994| 0.000205994| 0.02%|(func(), func2())
(call)| 1| 1.00112| 1.00112| 99.54%|# demo/threads.py:4 func
(call)| 1| 3.09944e-05| 3.09944e-05| 0.00%|# demo/threads.py:7 func2
15| 1| 7.62939e-05| 7.62939e-05| 0.01%|t1.join()
(call)| 1| 0.000423908| 0.000423908| 0.04%|# /usr/lib/python2.7/threading.py:653 join
16| 1| 5.26905e-05| 5.26905e-05| 0.01%|t2.join()
(call)| 1| 0.000320196| 0.000320196| 0.03%|# /usr/lib/python2.7/threading.py:653 join
Note that time.sleep call is not counted as such. For some reason, python is not generating c_call/c_return/c_exception events (which are ignored by current code, as a result).
Statistic profiling
In statistic profiling mode, pprofile periodically snapshots the current callstack(s) of current process to see what is being executed. As a result, profiler overhead can be dramatically reduced, making it possible to profile real workloads. Also, as statistic profiling acts at the whole-process level, it can be toggled independently of profiled code.
The downside of statistic profiling is that output lacks timing information, which makes it harder to understand.
$ pprofile --statistic .01 demo/threads.py
Command line: ['demo/threads.py']
Total duration: 1.0026s
File: demo/threads.py
File duration: 0s (0.00%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
1| 0| 0| 0| 0.00%|import threading
2| 0| 0| 0| 0.00%|import time
3| 0| 0| 0| 0.00%|
4| 0| 0| 0| 0.00%|def func():
5| 288| 0| 0| 0.00%| time.sleep(1)
6| 0| 0| 0| 0.00%|
7| 0| 0| 0| 0.00%|def func2():
8| 0| 0| 0| 0.00%| pass
9| 0| 0| 0| 0.00%|
10| 0| 0| 0| 0.00%|t1 = threading.Thread(target=func)
11| 0| 0| 0| 0.00%|t2 = threading.Thread(target=func)
12| 0| 0| 0| 0.00%|t1.start()
13| 0| 0| 0| 0.00%|t2.start()
14| 0| 0| 0| 0.00%|(func(), func2())
(call)| 96| 0| 0| 0.00%|# demo/threads.py:4 func
15| 0| 0| 0| 0.00%|t1.join()
16| 0| 0| 0| 0.00%|t2.join()
File: /usr/lib/python2.7/threading.py
File duration: 0s (0.00%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
[...]
308| 0| 0| 0| 0.00%| def wait(self, timeout=None):
[...]
338| 0| 0| 0| 0.00%| if timeout is None:
339| 1| 0| 0| 0.00%| waiter.acquire()
340| 0| 0| 0| 0.00%| if __debug__:
[...]
600| 0| 0| 0| 0.00%| def wait(self, timeout=None):
[...]
617| 0| 0| 0| 0.00%| if not self.__flag:
618| 0| 0| 0| 0.00%| self.__cond.wait(timeout)
(call)| 1| 0| 0| 0.00%|# /usr/lib/python2.7/threading.py:308 wait
[...]
724| 0| 0| 0| 0.00%| def start(self):
[...]
748| 0| 0| 0| 0.00%| self.__started.wait()
(call)| 1| 0| 0| 0.00%|# /usr/lib/python2.7/threading.py:600 wait
749| 0| 0| 0| 0.00%|
750| 0| 0| 0| 0.00%| def run(self):
[...]
760| 0| 0| 0| 0.00%| if self.__target:
761| 0| 0| 0| 0.00%| self.__target(*self.__args, **self.__kwargs)
(call)| 192| 0| 0| 0.00%|# demo/threads.py:4 func
762| 0| 0| 0| 0.00%| finally:
[...]
767| 0| 0| 0| 0.00%| def __bootstrap(self):
[...]
780| 0| 0| 0| 0.00%| try:
781| 0| 0| 0| 0.00%| self.__bootstrap_inner()
(call)| 192| 0| 0| 0.00%|# /usr/lib/python2.7/threading.py:790 __bootstrap_inner
[...]
790| 0| 0| 0| 0.00%| def __bootstrap_inner(self):
[...]
807| 0| 0| 0| 0.00%| try:
808| 0| 0| 0| 0.00%| self.run()
(call)| 192| 0| 0| 0.00%|# /usr/lib/python2.7/threading.py:750 run
Some details are lost (not all executed lines have a non-null hit-count), but the hot spot is still easily identifiable in this trivial example, and its call stack is still visible.
Thread-aware profiling
ThreadProfile class provides the same features as Profile, but uses
threading.settrace to propagate tracing to threading.Thread threads
started after profiling is enabled.
Limitations
The time spent in another thread is not discounted from interrupted line. On the long run, it should not be a problem if switches are evenly distributed among lines, but threads executing fewer lines will appear as eating more CPU time than they really do.
This is not specific to simultaneous multi-thread profiling: profiling a single thread of a multi-threaded application will also be polluted by time spent in other threads.
Example (lines are reported as taking longer to execute when profiled along with another thread - although the other thread is not profiled):
$ demo/embedded.py
Total duration: 1.00013s
File: demo/embedded.py
File duration: 1.00003s (99.99%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
1| 0| 0| 0| 0.00%|#!/usr/bin/env python
2| 0| 0| 0| 0.00%|import threading
3| 0| 0| 0| 0.00%|import pprofile
4| 0| 0| 0| 0.00%|import time
5| 0| 0| 0| 0.00%|import sys
6| 0| 0| 0| 0.00%|
7| 1| 1.5974e-05| 1.5974e-05| 0.00%|def func():
8| 0| 0| 0| 0.00%| # Busy loop, so context switches happen
9| 1| 1.40667e-05| 1.40667e-05| 0.00%| end = time.time() + 1
10| 146604| 0.511392| 3.48826e-06| 51.13%| while time.time() < end:
11| 146603| 0.48861| 3.33288e-06| 48.85%| pass
12| 0| 0| 0| 0.00%|
13| 0| 0| 0| 0.00%|# Single-treaded run
14| 0| 0| 0| 0.00%|prof = pprofile.Profile()
15| 0| 0| 0| 0.00%|with prof:
16| 0| 0| 0| 0.00%| func()
(call)| 1| 1.00003| 1.00003| 99.99%|# ./demo/embedded.py:7 func
17| 0| 0| 0| 0.00%|prof.annotate(sys.stdout, __file__)
18| 0| 0| 0| 0.00%|
19| 0| 0| 0| 0.00%|# Dual-threaded run
20| 0| 0| 0| 0.00%|t1 = threading.Thread(target=func)
21| 0| 0| 0| 0.00%|prof = pprofile.Profile()
22| 0| 0| 0| 0.00%|with prof:
23| 0| 0| 0| 0.00%| t1.start()
24| 0| 0| 0| 0.00%| func()
25| 0| 0| 0| 0.00%| t1.join()
26| 0| 0| 0| 0.00%|prof.annotate(sys.stdout, __file__)
Total duration: 1.00129s
File: demo/embedded.py
File duration: 1.00004s (99.88%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
[...]
7| 1| 1.50204e-05| 1.50204e-05| 0.00%|def func():
8| 0| 0| 0| 0.00%| # Busy loop, so context switches happen
9| 1| 2.38419e-05| 2.38419e-05| 0.00%| end = time.time() + 1
10| 64598| 0.538571| 8.33728e-06| 53.79%| while time.time() < end:
11| 64597| 0.461432| 7.14324e-06| 46.08%| pass
[...]
This also means that the sum of the percentage of all lines can exceed 100%. It can reach the number of concurrent threads (200% with 2 threads being busy for the whole profiled execution time, etc).
Example with 3 threads:
$ pprofile demo/threads.py
Command line: ['demo/threads.py']
Total duration: 1.00798s
File: demo/threads.py
File duration: 3.00604s (298.22%)
Line #| Hits| Time| Time per hit| %|Source code
------+----------+-------------+-------------+-------+-----------
1| 2| 3.21865e-05| 1.60933e-05| 0.00%|import threading
2| 1| 6.91414e-06| 6.91414e-06| 0.00%|import time
3| 0| 0| 0| 0.00%|
4| 4| 3.91006e-05| 9.77516e-06| 0.00%|def func():
5| 3| 3.00539| 1.0018|298.16%| time.sleep(1)
6| 0| 0| 0| 0.00%|
7| 2| 2.31266e-05| 1.15633e-05| 0.00%|def func2():
8| 1| 2.38419e-05| 2.38419e-05| 0.00%| pass
9| 0| 0| 0| 0.00%|
10| 1| 1.81198e-05| 1.81198e-05| 0.00%|t1 = threading.Thread(target=func)
(call)| 1| 0.000612974| 0.000612974| 0.06%|# /usr/lib/python2.7/threading.py:436 __init__
11| 1| 1.57356e-05| 1.57356e-05| 0.00%|t2 = threading.Thread(target=func)
(call)| 1| 0.000438213| 0.000438213| 0.04%|# /usr/lib/python2.7/threading.py:436 __init__
12| 1| 6.60419e-05| 6.60419e-05| 0.01%|t1.start()
(call)| 1| 0.000913858| 0.000913858| 0.09%|# /usr/lib/python2.7/threading.py:485 start
13| 1| 6.8903e-05| 6.8903e-05| 0.01%|t2.start()
(call)| 1| 0.00167513| 0.00167513| 0.17%|# /usr/lib/python2.7/threading.py:485 start
14| 1| 0.000200272| 0.000200272| 0.02%|(func(), func2())
(call)| 1| 1.00274| 1.00274| 99.48%|# demo/threads.py:4 func
(call)| 1| 4.19617e-05| 4.19617e-05| 0.00%|# demo/threads.py:7 func2
15| 1| 9.58443e-05| 9.58443e-05| 0.01%|t1.join()
(call)| 1| 0.000411987| 0.000411987| 0.04%|# /usr/lib/python2.7/threading.py:653 join
16| 1| 5.29289e-05| 5.29289e-05| 0.01%|t2.join()
(call)| 1| 0.000316143| 0.000316143| 0.03%|# /usr/lib/python2.7/threading.py:653 join
Note that the call time is not added to file total: it's already accounted for inside "func".
Why another profiler ?
Python's standard profiling tools have a callable-level granularity, which means it is only possible to tell which function is a hot-spot, not which lines in that function.
Robert Kern's line_profiler is a very nice alternative providing line-level profiling granularity, but in my opinion it has a few drawbacks which (in addition to the attractive technical challenge) made me start pprofile:
- It is not pure-python. This choice makes sense for performance but makes usage with pypy difficult and requires installation (I value execution straight from checkout).
- It requires source code modification to select what should be profiled. I prefer to have the option to do an in-depth, non-intrusive profiling.
- As an effect of previous point, it does not have a notion above individual callable, annotating functions but not whole files - preventing module import profiling.
- Profiling recursive code provides unexpected results (recursion cost is accumulated on callable's first line) because it doesn't track call stack. This may be unintended, and may be fixed at some point in line_profiler.