python - memory not being given back to kernel

Question

I have a very simple script that allocates memory, dels the only reference to a sizable object, all the while printing heapy and pidstat reports. After running the script, heapy tells me that there should not be much memory being used while pidstat tells me the opposite:

from guppy import hpy
import time
import sys
import os

'''
1) print heapy and pidstat report after starting and before actually doing any work
2) allocate some memory in a simple 2d array
3) print heapy and pidstat report
4) del the d2 array (attempt at garbage collection)
5) print heapy and pidstat report
6) sleep so pidstat can continue to be run to check on memory
'''

def pidstat(msg):
    print '==============================='
    print msg
    os.system('pidstat -r -p %s' % os.getpid())
    print '+++++++++++++++++++++++++++++++'
    print hpy().heap()[0]
    print '==============================='

pidstat('before doing anything')
docs = []
for doc in range(0, 10000):
    docs.append([j for j in range(0, 1000)])

pidstat('after fetching all the docs into memory')
del docs

pidstat('after freeing the docs')
time.sleep(60)

The output looks as follows:

===============================
before doing anything
Linux 2.6.38-15-generic (hersheezy)     08/14/2012  _x86_64_    (4 CPU)

01:05:20 PM       PID  minflt/s  majflt/s     VSZ    RSS   %MEM  Command
01:05:20 PM      5360      0.44      0.00   44768   9180   0.11  python
+++++++++++++++++++++++++++++++
Partition of a set of 19760 objects. Total size = 1591024 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0  19760 100  1591024 100   1591024 100 str
===============================
===============================
after fetching all the docs into memory
Linux 2.6.38-15-generic (hersheezy)     08/14/2012  _x86_64_    (4 CPU)

01:05:21 PM       PID  minflt/s  majflt/s     VSZ    RSS   %MEM  Command
01:05:21 PM      5360      8.95      0.00  318656 279120   3.49  python
+++++++++++++++++++++++++++++++
Partition of a set of 7431665 objects. Total size = 178359960 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0 7431665 100 178359960 100 178359960 100 int
===============================
===============================
after freeing the docs
Linux 2.6.38-15-generic (hersheezy)     08/14/2012  _x86_64_    (4 CPU)

01:05:29 PM       PID  minflt/s  majflt/s     VSZ    RSS   %MEM  Command
01:05:29 PM      5360     40.23      0.00  499984 460480   5.77  python
+++++++++++++++++++++++++++++++
Partition of a set of 19599 objects. Total size = 1582016 bytes.
 Index  Count   %     Size   % Cumulative  % Kind (class / dict of class)
     0  19599 100  1582016 100   1582016 100 str
===============================

How can I make sure this memory is returned to the operating system?

Answer 1

There can be a difference between when memory is made available for reuse inside the python process and when it is released to the OS. In particular, the standard Python interpreter (CPython) maintains its own pools and free lists for particular kinds of objects. It will reuse memory in these pools itself, but never releases it to the OS once it's been used.

See this for more details.

Answer 2

How can I make sure this memory is returned to the operating system?

It generally won't. Python allocates memory in 'arenas', and even when references are deleted in the interpreter, it will hold onto that memory arena to use later. I THINK there is a mechanism in newer version of python to unclaim arenas if they're completely empty. But you have no control over where your objects get placed.