如何剖析Python中的内存使用情况?

问题:如何剖析Python中的内存使用情况?

最近,我对算法产生了兴趣,并通过编写一个简单的实现,然后以各种方式对其进行了优化来开始探索它们。

我已经熟悉了用于分析运行时的标准Python模块(对于大多数事情,我发现IPython中的timeit magic函数就足够了),但是我也对内存使用感兴趣,因此我也可以探索这些折衷方案(例如,缓存先前计算的值与根据需要重新计算它们的表的成本)。是否有一个模块可以为我配置给定功能的内存使用情况?

I’ve recently become interested in algorithms and have begun exploring them by writing a naive implementation and then optimizing it in various ways.

I’m already familiar with the standard Python module for profiling runtime (for most things I’ve found the timeit magic function in IPython to be sufficient), but I’m also interested in memory usage so I can explore those tradeoffs as well (e.g. the cost of caching a table of previously computed values versus recomputing them as needed). Is there a module that will profile the memory usage of a given function for me?


回答 0

在这里已经回答了这个问题:Python memory profiler

基本上,您可以执行以下操作(引用自Guppy-PE):

>>> from guppy import hpy; h=hpy()
>>> h.heap()
Partition of a set of 48477 objects. Total size = 3265516 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0  25773  53  1612820  49   1612820  49 str
     1  11699  24   483960  15   2096780  64 tuple
     2    174   0   241584   7   2338364  72 dict of module
     3   3478   7   222592   7   2560956  78 types.CodeType
     4   3296   7   184576   6   2745532  84 function
     5    401   1   175112   5   2920644  89 dict of class
     6    108   0    81888   3   3002532  92 dict (no owner)
     7    114   0    79632   2   3082164  94 dict of type
     8    117   0    51336   2   3133500  96 type
     9    667   1    24012   1   3157512  97 __builtin__.wrapper_descriptor
<76 more rows. Type e.g. '_.more' to view.>
>>> h.iso(1,[],{})
Partition of a set of 3 objects. Total size = 176 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0      1  33      136  77       136  77 dict (no owner)
     1      1  33       28  16       164  93 list
     2      1  33       12   7       176 100 int
>>> x=[]
>>> h.iso(x).sp
 0: h.Root.i0_modules['__main__'].__dict__['x']
>>> 

This one has been answered already here: Python memory profiler

Basically you do something like that (cited from Guppy-PE):

>>> from guppy import hpy; h=hpy()
>>> h.heap()
Partition of a set of 48477 objects. Total size = 3265516 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0  25773  53  1612820  49   1612820  49 str
     1  11699  24   483960  15   2096780  64 tuple
     2    174   0   241584   7   2338364  72 dict of module
     3   3478   7   222592   7   2560956  78 types.CodeType
     4   3296   7   184576   6   2745532  84 function
     5    401   1   175112   5   2920644  89 dict of class
     6    108   0    81888   3   3002532  92 dict (no owner)
     7    114   0    79632   2   3082164  94 dict of type
     8    117   0    51336   2   3133500  96 type
     9    667   1    24012   1   3157512  97 __builtin__.wrapper_descriptor
<76 more rows. Type e.g. '_.more' to view.>
>>> h.iso(1,[],{})
Partition of a set of 3 objects. Total size = 176 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0      1  33      136  77       136  77 dict (no owner)
     1      1  33       28  16       164  93 list
     2      1  33       12   7       176 100 int
>>> x=[]
>>> h.iso(x).sp
 0: h.Root.i0_modules['__main__'].__dict__['x']
>>> 

回答 1

Python 3.4包含一个新模块:tracemalloc。它提供有关哪些代码分配最多内存的详细统计信息。这是显示分配内存的前三行的示例。

from collections import Counter
import linecache
import os
import tracemalloc

def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


tracemalloc.start()

counts = Counter()
fname = '/usr/share/dict/american-english'
with open(fname) as words:
    words = list(words)
    for word in words:
        prefix = word[:3]
        counts[prefix] += 1
print('Top prefixes:', counts.most_common(3))

snapshot = tracemalloc.take_snapshot()
display_top(snapshot)

结果如下:

Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
Top 3 lines
#1: scratches/memory_test.py:37: 6527.1 KiB
    words = list(words)
#2: scratches/memory_test.py:39: 247.7 KiB
    prefix = word[:3]
#3: scratches/memory_test.py:40: 193.0 KiB
    counts[prefix] += 1
4 other: 4.3 KiB
Total allocated size: 6972.1 KiB

什么时候内存泄漏不是泄漏?

当计算结束时仍保留内存时,该示例非常有用,但是有时您拥有分配大量内存然后释放所有内存的代码。从技术上讲,这不是内存泄漏,但是它使用的内存比您想象的要多。释放所有内存时如何跟踪?如果是您的代码,则可能可以添加一些调试代码以在运行时拍摄快照。如果没有,您可以在主线程运行时启动后台线程来监视内存使用情况。

这是前面的示例,其中所有代码都已移入count_prefixes()函数中。该函数返回时,将释放所有内存。我还添加了一些sleep()调用来模拟长时间运行的计算。

from collections import Counter
import linecache
import os
import tracemalloc
from time import sleep


def count_prefixes():
    sleep(2)  # Start up time.
    counts = Counter()
    fname = '/usr/share/dict/american-english'
    with open(fname) as words:
        words = list(words)
        for word in words:
            prefix = word[:3]
            counts[prefix] += 1
            sleep(0.0001)
    most_common = counts.most_common(3)
    sleep(3)  # Shut down time.
    return most_common


def main():
    tracemalloc.start()

    most_common = count_prefixes()
    print('Top prefixes:', most_common)

    snapshot = tracemalloc.take_snapshot()
    display_top(snapshot)


def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


main()

当我运行该版本时,内存使用已从6MB减少到4KB,因为该函数在完成时会释放其所有内存。

Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
Top 3 lines
#1: collections/__init__.py:537: 0.7 KiB
    self.update(*args, **kwds)
#2: collections/__init__.py:555: 0.6 KiB
    return _heapq.nlargest(n, self.items(), key=_itemgetter(1))
#3: python3.6/heapq.py:569: 0.5 KiB
    result = [(key(elem), i, elem) for i, elem in zip(range(0, -n, -1), it)]
10 other: 2.2 KiB
Total allocated size: 4.0 KiB

现在,这是受另一个答案启发的版本,该答案启动了另一个线程来监视内存使用情况。

from collections import Counter
import linecache
import os
import tracemalloc
from datetime import datetime
from queue import Queue, Empty
from resource import getrusage, RUSAGE_SELF
from threading import Thread
from time import sleep

def memory_monitor(command_queue: Queue, poll_interval=1):
    tracemalloc.start()
    old_max = 0
    snapshot = None
    while True:
        try:
            command_queue.get(timeout=poll_interval)
            if snapshot is not None:
                print(datetime.now())
                display_top(snapshot)

            return
        except Empty:
            max_rss = getrusage(RUSAGE_SELF).ru_maxrss
            if max_rss > old_max:
                old_max = max_rss
                snapshot = tracemalloc.take_snapshot()
                print(datetime.now(), 'max RSS', max_rss)


def count_prefixes():
    sleep(2)  # Start up time.
    counts = Counter()
    fname = '/usr/share/dict/american-english'
    with open(fname) as words:
        words = list(words)
        for word in words:
            prefix = word[:3]
            counts[prefix] += 1
            sleep(0.0001)
    most_common = counts.most_common(3)
    sleep(3)  # Shut down time.
    return most_common


def main():
    queue = Queue()
    poll_interval = 0.1
    monitor_thread = Thread(target=memory_monitor, args=(queue, poll_interval))
    monitor_thread.start()
    try:
        most_common = count_prefixes()
        print('Top prefixes:', most_common)
    finally:
        queue.put('stop')
        monitor_thread.join()


def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


main()

resource模块使您可以检查当前内存使用情况,并从峰值内存使用情况中保存快照。队列让主线程告诉内存监视器线程何时打印其报告并关闭。运行时,它显示list()调用正在使用的内存:

2018-05-29 10:34:34.441334 max RSS 10188
2018-05-29 10:34:36.475707 max RSS 23588
2018-05-29 10:34:36.616524 max RSS 38104
2018-05-29 10:34:36.772978 max RSS 45924
2018-05-29 10:34:36.929688 max RSS 46824
2018-05-29 10:34:37.087554 max RSS 46852
Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
2018-05-29 10:34:56.281262
Top 3 lines
#1: scratches/scratch.py:36: 6527.0 KiB
    words = list(words)
#2: scratches/scratch.py:38: 16.4 KiB
    prefix = word[:3]
#3: scratches/scratch.py:39: 10.1 KiB
    counts[prefix] += 1
19 other: 10.8 KiB
Total allocated size: 6564.3 KiB

如果您使用的是Linux,则可能会发现/proc/self/statm比该resource模块更有用。

Python 3.4 includes a new module: tracemalloc. It provides detailed statistics about which code is allocating the most memory. Here’s an example that displays the top three lines allocating memory.

from collections import Counter
import linecache
import os
import tracemalloc

def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


tracemalloc.start()

counts = Counter()
fname = '/usr/share/dict/american-english'
with open(fname) as words:
    words = list(words)
    for word in words:
        prefix = word[:3]
        counts[prefix] += 1
print('Top prefixes:', counts.most_common(3))

snapshot = tracemalloc.take_snapshot()
display_top(snapshot)

And here are the results:

Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
Top 3 lines
#1: scratches/memory_test.py:37: 6527.1 KiB
    words = list(words)
#2: scratches/memory_test.py:39: 247.7 KiB
    prefix = word[:3]
#3: scratches/memory_test.py:40: 193.0 KiB
    counts[prefix] += 1
4 other: 4.3 KiB
Total allocated size: 6972.1 KiB

When is a memory leak not a leak?

That example is great when the memory is still being held at the end of the calculation, but sometimes you have code that allocates a lot of memory and then releases it all. It’s not technically a memory leak, but it’s using more memory than you think it should. How can you track memory usage when it all gets released? If it’s your code, you can probably add some debugging code to take snapshots while it’s running. If not, you can start a background thread to monitor memory usage while the main thread runs.

Here’s the previous example where the code has all been moved into the count_prefixes() function. When that function returns, all the memory is released. I also added some sleep() calls to simulate a long-running calculation.

from collections import Counter
import linecache
import os
import tracemalloc
from time import sleep


def count_prefixes():
    sleep(2)  # Start up time.
    counts = Counter()
    fname = '/usr/share/dict/american-english'
    with open(fname) as words:
        words = list(words)
        for word in words:
            prefix = word[:3]
            counts[prefix] += 1
            sleep(0.0001)
    most_common = counts.most_common(3)
    sleep(3)  # Shut down time.
    return most_common


def main():
    tracemalloc.start()

    most_common = count_prefixes()
    print('Top prefixes:', most_common)

    snapshot = tracemalloc.take_snapshot()
    display_top(snapshot)


def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


main()

When I run that version, the memory usage has gone from 6MB down to 4KB, because the function released all its memory when it finished.

Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
Top 3 lines
#1: collections/__init__.py:537: 0.7 KiB
    self.update(*args, **kwds)
#2: collections/__init__.py:555: 0.6 KiB
    return _heapq.nlargest(n, self.items(), key=_itemgetter(1))
#3: python3.6/heapq.py:569: 0.5 KiB
    result = [(key(elem), i, elem) for i, elem in zip(range(0, -n, -1), it)]
10 other: 2.2 KiB
Total allocated size: 4.0 KiB

Now here’s a version inspired by another answer that starts a second thread to monitor memory usage.

from collections import Counter
import linecache
import os
import tracemalloc
from datetime import datetime
from queue import Queue, Empty
from resource import getrusage, RUSAGE_SELF
from threading import Thread
from time import sleep

def memory_monitor(command_queue: Queue, poll_interval=1):
    tracemalloc.start()
    old_max = 0
    snapshot = None
    while True:
        try:
            command_queue.get(timeout=poll_interval)
            if snapshot is not None:
                print(datetime.now())
                display_top(snapshot)

            return
        except Empty:
            max_rss = getrusage(RUSAGE_SELF).ru_maxrss
            if max_rss > old_max:
                old_max = max_rss
                snapshot = tracemalloc.take_snapshot()
                print(datetime.now(), 'max RSS', max_rss)


def count_prefixes():
    sleep(2)  # Start up time.
    counts = Counter()
    fname = '/usr/share/dict/american-english'
    with open(fname) as words:
        words = list(words)
        for word in words:
            prefix = word[:3]
            counts[prefix] += 1
            sleep(0.0001)
    most_common = counts.most_common(3)
    sleep(3)  # Shut down time.
    return most_common


def main():
    queue = Queue()
    poll_interval = 0.1
    monitor_thread = Thread(target=memory_monitor, args=(queue, poll_interval))
    monitor_thread.start()
    try:
        most_common = count_prefixes()
        print('Top prefixes:', most_common)
    finally:
        queue.put('stop')
        monitor_thread.join()


def display_top(snapshot, key_type='lineno', limit=3):
    snapshot = snapshot.filter_traces((
        tracemalloc.Filter(False, "<frozen importlib._bootstrap>"),
        tracemalloc.Filter(False, "<unknown>"),
    ))
    top_stats = snapshot.statistics(key_type)

    print("Top %s lines" % limit)
    for index, stat in enumerate(top_stats[:limit], 1):
        frame = stat.traceback[0]
        # replace "/path/to/module/file.py" with "module/file.py"
        filename = os.sep.join(frame.filename.split(os.sep)[-2:])
        print("#%s: %s:%s: %.1f KiB"
              % (index, filename, frame.lineno, stat.size / 1024))
        line = linecache.getline(frame.filename, frame.lineno).strip()
        if line:
            print('    %s' % line)

    other = top_stats[limit:]
    if other:
        size = sum(stat.size for stat in other)
        print("%s other: %.1f KiB" % (len(other), size / 1024))
    total = sum(stat.size for stat in top_stats)
    print("Total allocated size: %.1f KiB" % (total / 1024))


main()

The resource module lets you check the current memory usage, and save the snapshot from the peak memory usage. The queue lets the main thread tell the memory monitor thread when to print its report and shut down. When it runs, it shows the memory being used by the list() call:

2018-05-29 10:34:34.441334 max RSS 10188
2018-05-29 10:34:36.475707 max RSS 23588
2018-05-29 10:34:36.616524 max RSS 38104
2018-05-29 10:34:36.772978 max RSS 45924
2018-05-29 10:34:36.929688 max RSS 46824
2018-05-29 10:34:37.087554 max RSS 46852
Top prefixes: [('con', 1220), ('dis', 1002), ('pro', 809)]
2018-05-29 10:34:56.281262
Top 3 lines
#1: scratches/scratch.py:36: 6527.0 KiB
    words = list(words)
#2: scratches/scratch.py:38: 16.4 KiB
    prefix = word[:3]
#3: scratches/scratch.py:39: 10.1 KiB
    counts[prefix] += 1
19 other: 10.8 KiB
Total allocated size: 6564.3 KiB

If you’re on Linux, you may find /proc/self/statm more useful than the resource module.


回答 2

如果只想查看对象的内存使用情况,(回答其他问题

有一个名为Pymplerasizeof 模块,其中包含该模块。

用法如下:

from pympler import asizeof
asizeof.asizeof(my_object)

sys.getsizeof与之不同,它适用于您自己创建的对象

>>> asizeof.asizeof(tuple('bcd'))
200
>>> asizeof.asizeof({'foo': 'bar', 'baz': 'bar'})
400
>>> asizeof.asizeof({})
280
>>> asizeof.asizeof({'foo':'bar'})
360
>>> asizeof.asizeof('foo')
40
>>> asizeof.asizeof(Bar())
352
>>> asizeof.asizeof(Bar().__dict__)
280
>>> help(asizeof.asizeof)
Help on function asizeof in module pympler.asizeof:

asizeof(*objs, **opts)
    Return the combined size in bytes of all objects passed as positional arguments.

If you only want to look at the memory usage of an object, (answer to other question)

There is a module called Pympler which contains the asizeof module.

Use as follows:

from pympler import asizeof
asizeof.asizeof(my_object)

Unlike sys.getsizeof, it works for your self-created objects.

>>> asizeof.asizeof(tuple('bcd'))
200
>>> asizeof.asizeof({'foo': 'bar', 'baz': 'bar'})
400
>>> asizeof.asizeof({})
280
>>> asizeof.asizeof({'foo':'bar'})
360
>>> asizeof.asizeof('foo')
40
>>> asizeof.asizeof(Bar())
352
>>> asizeof.asizeof(Bar().__dict__)
280
>>> help(asizeof.asizeof)
Help on function asizeof in module pympler.asizeof:

asizeof(*objs, **opts)
    Return the combined size in bytes of all objects passed as positional arguments.

回答 3

披露:

  • 仅适用于Linux
  • 报告用于由当前过程作为一个整体,而不是单个存储器功能

但由于它的简单性,它很不错:

import resource
def using(point=""):
    usage=resource.getrusage(resource.RUSAGE_SELF)
    return '''%s: usertime=%s systime=%s mem=%s mb
           '''%(point,usage[0],usage[1],
                usage[2]/1024.0 )

只需插入using("Label")您想查看的情况即可。例如

print(using("before"))
wrk = ["wasting mem"] * 1000000
print(using("after"))

>>> before: usertime=2.117053 systime=1.703466 mem=53.97265625 mb
>>> after: usertime=2.12023 systime=1.70708 mem=60.8828125 mb

Disclosure:

  • Applicable on Linux only
  • Reports memory used by the current process as a whole, not individual functions within

But nice because of its simplicity:

import resource
def using(point=""):
    usage=resource.getrusage(resource.RUSAGE_SELF)
    return '''%s: usertime=%s systime=%s mem=%s mb
           '''%(point,usage[0],usage[1],
                usage[2]/1024.0 )

Just insert using("Label") where you want to see what’s going on. For example

print(using("before"))
wrk = ["wasting mem"] * 1000000
print(using("after"))

>>> before: usertime=2.117053 systime=1.703466 mem=53.97265625 mb
>>> after: usertime=2.12023 systime=1.70708 mem=60.8828125 mb

回答 4

在我看来,既然已接受的答案以及投票数第二高的答案都存在一些问题,所以我想再提供一个基于Ihor B.答案的答案,并进行了一些微小但重要的修改。

该解决方案允许您运行分析上或者通过包装函数调用用profile,功能和调用它通过与装饰你的函数/法@profile装饰。

当您要分析一些第三方代码而不弄乱其源代码时,第一种技术很有用,而第二种技术则比较“干净”,当您不介意修改函数/方法的源代码时,效果更好想要简介。

我还修改了输出,以便获得RSS,VMS和共享内存。我不太关心“之前”和“之后”的值,只关心增量,所以我删除了那些值(如果您要与Ihor B.的答案进行比较)。

分析代码

# profile.py
import time
import os
import psutil
import inspect


def elapsed_since(start):
    #return time.strftime("%H:%M:%S", time.gmtime(time.time() - start))
    elapsed = time.time() - start
    if elapsed < 1:
        return str(round(elapsed*1000,2)) + "ms"
    if elapsed < 60:
        return str(round(elapsed, 2)) + "s"
    if elapsed < 3600:
        return str(round(elapsed/60, 2)) + "min"
    else:
        return str(round(elapsed / 3600, 2)) + "hrs"


def get_process_memory():
    process = psutil.Process(os.getpid())
    mi = process.memory_info()
    return mi.rss, mi.vms, mi.shared


def format_bytes(bytes):
    if abs(bytes) < 1000:
        return str(bytes)+"B"
    elif abs(bytes) < 1e6:
        return str(round(bytes/1e3,2)) + "kB"
    elif abs(bytes) < 1e9:
        return str(round(bytes / 1e6, 2)) + "MB"
    else:
        return str(round(bytes / 1e9, 2)) + "GB"


def profile(func, *args, **kwargs):
    def wrapper(*args, **kwargs):
        rss_before, vms_before, shared_before = get_process_memory()
        start = time.time()
        result = func(*args, **kwargs)
        elapsed_time = elapsed_since(start)
        rss_after, vms_after, shared_after = get_process_memory()
        print("Profiling: {:>20}  RSS: {:>8} | VMS: {:>8} | SHR {"
              ":>8} | time: {:>8}"
            .format("<" + func.__name__ + ">",
                    format_bytes(rss_after - rss_before),
                    format_bytes(vms_after - vms_before),
                    format_bytes(shared_after - shared_before),
                    elapsed_time))
        return result
    if inspect.isfunction(func):
        return wrapper
    elif inspect.ismethod(func):
        return wrapper(*args,**kwargs)

用法示例,假设上面的代码另存为profile.py

from profile import profile
from time import sleep
from sklearn import datasets # Just an example of 3rd party function call


# Method 1
run_profiling = profile(datasets.load_digits)
data = run_profiling()

# Method 2
@profile
def my_function():
    # do some stuff
    a_list = []
    for i in range(1,100000):
        a_list.append(i)
    return a_list


res = my_function()

这将导致输出类似于以下内容:

Profiling:        <load_digits>  RSS:   5.07MB | VMS:   4.91MB | SHR  73.73kB | time:  89.99ms
Profiling:        <my_function>  RSS:   1.06MB | VMS:   1.35MB | SHR       0B | time:   8.43ms

重要的最后几点注意事项:

  1. 请记住,这种剖析方法仅是近似的,因为计算机上可能会发生许多其他事情。由于垃圾收集和其他因素,增量甚至可能为零。
  2. 由于某些未知的原因,出现非常短的函数调用(例如1或2 ms),而内存使用量为零。我怀疑这是硬件/操作系统(在装有Linux的基本笔记本电脑上测试过)在内存统计信息更新频率方面的一些限制。
  3. 为了使示例简单,我没有使用任何函数参数,但是它们应该像预期的那样工作,即 profile(my_function, arg)进行概要分析my_function(arg)

Since the accepted answer and also the next highest voted answer have, in my opinion, some problems, I’d like to offer one more answer that is based closely on Ihor B.’s answer with some small but important modifications.

This solution allows you to run profiling on either by wrapping a function call with the profile function and calling it, or by decorating your function/method with the @profile decorator.

The first technique is useful when you want to profile some third-party code without messing with its source, whereas the second technique is a bit “cleaner” and works better when you are don’t mind modifying the source of the function/method you want to profile.

I’ve also modified the output, so that you get RSS, VMS, and shared memory. I don’t care much about the “before” and “after” values, but only the delta, so I removed those (if you’re comparing to Ihor B.’s answer).

Profiling code

# profile.py
import time
import os
import psutil
import inspect


def elapsed_since(start):
    #return time.strftime("%H:%M:%S", time.gmtime(time.time() - start))
    elapsed = time.time() - start
    if elapsed < 1:
        return str(round(elapsed*1000,2)) + "ms"
    if elapsed < 60:
        return str(round(elapsed, 2)) + "s"
    if elapsed < 3600:
        return str(round(elapsed/60, 2)) + "min"
    else:
        return str(round(elapsed / 3600, 2)) + "hrs"


def get_process_memory():
    process = psutil.Process(os.getpid())
    mi = process.memory_info()
    return mi.rss, mi.vms, mi.shared


def format_bytes(bytes):
    if abs(bytes) < 1000:
        return str(bytes)+"B"
    elif abs(bytes) < 1e6:
        return str(round(bytes/1e3,2)) + "kB"
    elif abs(bytes) < 1e9:
        return str(round(bytes / 1e6, 2)) + "MB"
    else:
        return str(round(bytes / 1e9, 2)) + "GB"


def profile(func, *args, **kwargs):
    def wrapper(*args, **kwargs):
        rss_before, vms_before, shared_before = get_process_memory()
        start = time.time()
        result = func(*args, **kwargs)
        elapsed_time = elapsed_since(start)
        rss_after, vms_after, shared_after = get_process_memory()
        print("Profiling: {:>20}  RSS: {:>8} | VMS: {:>8} | SHR {"
              ":>8} | time: {:>8}"
            .format("<" + func.__name__ + ">",
                    format_bytes(rss_after - rss_before),
                    format_bytes(vms_after - vms_before),
                    format_bytes(shared_after - shared_before),
                    elapsed_time))
        return result
    if inspect.isfunction(func):
        return wrapper
    elif inspect.ismethod(func):
        return wrapper(*args,**kwargs)

Example usage, assuming the above code is saved as profile.py:

from profile import profile
from time import sleep
from sklearn import datasets # Just an example of 3rd party function call


# Method 1
run_profiling = profile(datasets.load_digits)
data = run_profiling()

# Method 2
@profile
def my_function():
    # do some stuff
    a_list = []
    for i in range(1,100000):
        a_list.append(i)
    return a_list


res = my_function()

This should result in output similar to the below:

Profiling:        <load_digits>  RSS:   5.07MB | VMS:   4.91MB | SHR  73.73kB | time:  89.99ms
Profiling:        <my_function>  RSS:   1.06MB | VMS:   1.35MB | SHR       0B | time:   8.43ms

A couple of important final notes:

  1. Keep in mind, this method of profiling is only going to be approximate, since lots of other stuff might be happening on the machine. Due to garbage collection and other factors, the deltas might even be zero.
  2. For some unknown reason, very short function calls (e.g. 1 or 2 ms) show up with zero memory usage. I suspect this is some limitation of the hardware/OS (tested on basic laptop with Linux) on how often memory statistics are updated.
  3. To keep the examples simple, I didn’t use any function arguments, but they should work as one would expect, i.e. profile(my_function, arg) to profile my_function(arg)

回答 5

下面是一个简单的函数装饰器,它可以跟踪函数调用之前,函数调用之后进程消耗的内存量以及它们之间的区别:

import time
import os
import psutil


def elapsed_since(start):
    return time.strftime("%H:%M:%S", time.gmtime(time.time() - start))


def get_process_memory():
    process = psutil.Process(os.getpid())
    return process.get_memory_info().rss


def profile(func):
    def wrapper(*args, **kwargs):
        mem_before = get_process_memory()
        start = time.time()
        result = func(*args, **kwargs)
        elapsed_time = elapsed_since(start)
        mem_after = get_process_memory()
        print("{}: memory before: {:,}, after: {:,}, consumed: {:,}; exec time: {}".format(
            func.__name__,
            mem_before, mem_after, mem_after - mem_before,
            elapsed_time))
        return result
    return wrapper

这是我的博客,描述了所有详细信息。(已归档的链接

Below is a simple function decorator which allows to track how much memory the process consumed before the function call, after the function call, and what is the difference:

import time
import os
import psutil
 
 
def elapsed_since(start):
    return time.strftime("%H:%M:%S", time.gmtime(time.time() - start))
 
 
def get_process_memory():
    process = psutil.Process(os.getpid())
    mem_info = process.memory_info()
    return mem_info.rss
 
 
def profile(func):
    def wrapper(*args, **kwargs):
        mem_before = get_process_memory()
        start = time.time()
        result = func(*args, **kwargs)
        elapsed_time = elapsed_since(start)
        mem_after = get_process_memory()
        print("{}: memory before: {:,}, after: {:,}, consumed: {:,}; exec time: {}".format(
            func.__name__,
            mem_before, mem_after, mem_after - mem_before,
            elapsed_time))
        return result
    return wrapper

Here is my blog which describes all the details. (archived link)


回答 6

也许有帮助:
< 参见其他 >

pip install gprof2dot
sudo apt-get install graphviz

gprof2dot -f pstats profile_for_func1_001 | dot -Tpng -o profile.png

def profileit(name):
    """
    @profileit("profile_for_func1_001")
    """
    def inner(func):
        def wrapper(*args, **kwargs):
            prof = cProfile.Profile()
            retval = prof.runcall(func, *args, **kwargs)
            # Note use of name from outer scope
            prof.dump_stats(name)
            return retval
        return wrapper
    return inner

@profileit("profile_for_func1_001")
def func1(...)

maybe it help:
<see additional>

pip install gprof2dot
sudo apt-get install graphviz

gprof2dot -f pstats profile_for_func1_001 | dot -Tpng -o profile.png

def profileit(name):
    """
    @profileit("profile_for_func1_001")
    """
    def inner(func):
        def wrapper(*args, **kwargs):
            prof = cProfile.Profile()
            retval = prof.runcall(func, *args, **kwargs)
            # Note use of name from outer scope
            prof.dump_stats(name)
            return retval
        return wrapper
    return inner

@profileit("profile_for_func1_001")
def func1(...)

回答 7

一个简单的示例,使用memory_profile计算代码块/函数的内存使用率,同时返回函数的结果:

import memory_profiler as mp

def fun(n):
    tmp = []
    for i in range(n):
        tmp.extend(list(range(i*i)))
    return "XXXXX"

在运行代码之前计算内存使用量,然后在代码执行期间计算最大使用量:

start_mem = mp.memory_usage(max_usage=True)
res = mp.memory_usage(proc=(fun, [100]), max_usage=True, retval=True) 
print('start mem', start_mem)
print('max mem', res[0][0])
print('used mem', res[0][0]-start_mem)
print('fun output', res[1])

计算运行功能时采样点的使用情况:

res = mp.memory_usage((fun, [100]), interval=.001, retval=True)
print('min mem', min(res[0]))
print('max mem', max(res[0]))
print('used mem', max(res[0])-min(res[0]))
print('fun output', res[1])

积分:@skeept

A simple example to calculate the memory usage of a block of codes / function using memory_profile, while returning result of the function:

import memory_profiler as mp

def fun(n):
    tmp = []
    for i in range(n):
        tmp.extend(list(range(i*i)))
    return "XXXXX"

calculate memory usage before running the code then calculate max usage during the code:

start_mem = mp.memory_usage(max_usage=True)
res = mp.memory_usage(proc=(fun, [100]), max_usage=True, retval=True) 
print('start mem', start_mem)
print('max mem', res[0][0])
print('used mem', res[0][0]-start_mem)
print('fun output', res[1])

calculate usage in sampling points while running function:

res = mp.memory_usage((fun, [100]), interval=.001, retval=True)
print('min mem', min(res[0]))
print('max mem', max(res[0]))
print('used mem', max(res[0])-min(res[0]))
print('fun output', res[1])

Credits: @skeept