Python and memory fragmentation

If you use CPython on 32 bit architectures, you may encounter a problem called memory fragmentation. It is more likely to happen on Windows for reasons that will soon be clear, but it's not a Windows exclusive. It is also not an exclusive python problem, but tends to occur more often on CPython due to its intrinsic memory allocation strategy.

When you dynamically allocate memory in C, you do so in a particular area of the Virtual Memory, the heap. A requirement for this allocation is that the allocated chunk must be contiguous in virtual memory. CPython puts extreme stress on the heap: all objects are allocated dynamically, and when they are freed, a hole is left in the heap. This hole may be filled by a later allocation, if the requested memory fits in the hole, but if it doesn't, the hole remains until something that fits is requested. On Linux, you can follow the VM occupation with this small python script

import sys
import subprocess

mmap = [' ']*(16*256)
out = subprocess.check_output(["/usr/bin/pmap","-x", "%d" % int(sys.argv[1])])
for i in out.splitlines()[2:-2]:
    values = i.split()[0:2]
    start = int("0x"+values[0], 16) / 2**20
    end = start + (int(values[1])*1024)/2**20
    for p in xrange(start, end+1):
        mmap[p] = '*'

for row in xrange(16):
    print hex(row)+" | "+"".join( 
                        mmap[row * 256:(row+1)*256]
                        )+"|"

On Windows, the great utility VMMap can be used to monitor the same information.

Given the above scenario, the Virtual Memory space can eventually become extremely fragmented, depending on the size of your objects, their order of allocation, if your application jumps between dynamically allocating large chunks of memory and small python objects, and so on. As a result, you may not be able to perform a large allocation, not because you are out of memory, but because you are out of contiguous memory in your VM address space. In a recent benchmark I performed on Windows 7, the largest contiguous chunk of memory available was a meager 32 megabytes (ow!), which means that despite the free memory being around 1 gigabyte, the biggest chunk I could request was only 32 megabytes. Anything bigger would have the malloc fail.

Additional conditions that can make the problem worse are dynamic libraries binding and excessive threading. The first invades your VM address space, and the second needs a stack per each thread, putting additional unmovable barriers throughout the VM and reducing real estate for contiguous blocks. See for example what happens with 10 threads on Linux

(gdb) thread apply all print $esp

Thread 10 (Thread 10606): $5 = (void *) 0x50d7184
Thread 9 (Thread 10607): $6 = (void *) 0x757ce90
Thread 8 (Thread 10608): $7 = (void *) 0x7b69e90
Thread 7 (Thread 10609): $8 = (void *) 0x7d6ae90
Thread 6 (Thread 10618): $9 = (void *) 0x8a4ae90
Thread 5 (Thread 10619): $10 = (void *) 0xb22fee90
Thread 4 (Thread 10620): $11 = (void *) 0xb20fde90
Thread 3 (Thread 10621): $12 = (void *) 0xb1efce90
Thread 2 (Thread 10806): $13 = (void *) 0xb2ea31c4
Thread 1 (Thread 0xb7f6b6c0 (LWP 10593)): $14 = (void *) 0xbffd1f3c

Fragmentation is eventually made irrelevant by a 64 bit architecture, where the VM address space is huge (for now ;) ). Yet, if you have a 32 bit machine and a long running python process that is juggling large and small allocations, you may eventually run out of contiguous memory and see malloc() fail.

How to solve? I found this interesting article that details the same issue and provides some mitigation techniques for Windows, because Windows is kind of special: on Windows, the 4 gigabytes address space is divided in two parts of 2GB EACH, the first for the process, the second reserved for the kernel. If you must stay on 32 bits, your best bet is to give an additional gigabyte of VM with the following recipe (only valid for Windows-7)

  1. run bcdedit /set IncreaseUserVa 3072 as administrator, then reboot the machine.
  2. mark your executable with EditBin.exe yourprogram.exe /LARGEADDRESSAWARE

With this command, the VM subdivision is set to 3GB+1GB, granting one additional gigabyte to your process. This improves the situation, but sometimes it's enough. If it is not, and you still need to work on 32 bit machines then you are in big trouble. You have to change your allocation strategy and be smarter on handling the fragmentation within your code.