如何在Python中分析内存使用情况？

Question

我最近对算法感兴趣,并开始通过编写一个简单的实现,然后以各种方式优化它来探索它们.

我已经熟悉用于分析运行时的标准Python模块(对于大多数事情我已经发现IPython中的timeit魔术功能已足够),但我也对内存使用感兴趣,所以我也可以探索这些权衡(例如,缓存先前计算的值表的成本与根据需要重新计算它们的成本.是否有一个模块可以为我分析给定函数的内存使用情况？

Answer 1

这个已经在这里得到了解答:Python内存分析器

基本上你做了类似的事情(引自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']
>>> 

Answer 2

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.这是一个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)

以下是结果:

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？

当计算结束时仍然保留内存时,这个例子很棒,但有时你会有分配大量内存的代码,然后释放所有内存.它在技术上不是内存泄漏,但它使用的内存比你想象的要多.如何释放内存时如何跟踪内存使用情况？如果它是您的代码,您可以添加一些调试代码以在其运行时拍摄快照.如果没有,您可以启动后台线程来监视主线程运行时的内存使用情况.

这是前面的示例,其中代码已全部移入count_prefixes()函数中.当该函数返回时,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,内存使用量从6MB减少到4KB,因为该函数在完成后释放了所有内存.

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,这是一个受另一个答案启发的版本,它启动了第二个线程来监控内存使用情况.

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()

该resource模块允许您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.队列允许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

如果你在Linux上,你会发现/proc/self/statm比resource模块更有用.

Answer 3

如果你只想查看一个对象的内存使用情况,(回答其他问题)

有一个名为Pympler的asizeof 模块,它包含模块.

使用方法如下:

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.

Answer 4

对于一个非常简单的方法尝试:

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]*resource.getpagesize())/1000000.0 )

只需插入using("Label")您想要查看正在发生的事情的位置即可.

Answer 5

一个使用 memory_profile 计算代码块/函数的内存使用情况的简单示例，同时返回函数结果：

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情况，然后计算代码期间的最大使用情况：

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])

计算运行函数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])

学分：@skeept

Answer 6

我认为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，并进行了一些小而重要的修改。

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，功能和调用它或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，因此我删除了这些值（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，这种分析方法仅是近似的，因为计算机上可能会发生很多其他事情。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)

Answer 7

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）