Failure to create multiprocessing shared arrays larger than 50% of memory size under linux

[mboquien]

BPO	21116
Nosy	@pitrou, @serhiy-storchaka
Files	shared_array.diff shared_array.diff

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2015-04-13.18:55:22.032>
created_at = <Date 2014-03-31.19:37:47.936>
labels = ['library', 'performance']
title = 'Failure to create multiprocessing shared arrays larger than 50% of memory size under linux'
updated_at = <Date 2015-04-14.20:58:22.873>
user = 'https://bugs.python.org/mboquien'

bugs.python.org fields:

activity = <Date 2015-04-14.20:58:22.873>
actor = 'pitrou'
assignee = 'none'
closed = True
closed_date = <Date 2015-04-13.18:55:22.032>
closer = 'pitrou'
components = ['Library (Lib)']
creation = <Date 2014-03-31.19:37:47.936>
creator = 'mboquien'
dependencies = []
files = ['34686', '34687']
hgrepos = []
issue_num = 21116
keywords = ['patch']
message_count = 18.0
messages = ['215258', '215260', '215264', '215268', '215296', '215404', '215407', '215408', '215425', '215433', '215460', '215481', '215494', '215583', '240704', '240705', '240874', '241027']
nosy_count = 6.0
nosy_names = ['pitrou', 'neologix', 'python-dev', 'sbt', 'serhiy.storchaka', 'mboquien']
pr_nums = []
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'resource usage'
url = 'https://bugs.python.org/issue21116'
versions = ['Python 3.5']

[mboquien]

It is currently impossible to create multiprocessing shared arrays larger than 50% of memory size under linux (and I assume other unices). A simple test case would be the following:

from multiprocessing.sharedctypes import RawArray
import ctypes

foo = RawArray(ctypes.c_double, 10*1024**3//8)  # Allocate 10GB array

If the array is larger than 50% of the total memory size, the process get SIGKILL'ed by the OS. Deactivate the swap for better effects.

Naturally this requires that the tmpfs max size is large enough, which is the case here, 15GB max with 16GB of RAM.

I have tracked down the problem to multiprocessing/heap.py. The guilty line is: f.write(b'\0'*size). Indeed, for very large sizes it is going to create a large intermediate array (10 GB in my test case) and as much memory is going to be allocated to the new shared array, leading to a memory consumption over the limit.

To solve the problem, I have split the zeroing of the shared array into blocks of 1MB. I can now allocate arrays as large as the tmpfs maximum size. Also it runs a bit faster. On a test case of a 6GB RawArray, 3.4.0 takes a total time of 3.930s whereas it goes down to 3.061s with the attached patch.

[mboquien]

Updated the patch not to create a uselessly large array if the size is small than the block size.

[mboquien]

New update of the patch following Antoine Pitrou's comments. PEP-8 does not complain anymore.

[pitrou]

You overlooked the part where I was suggesting to add a unit test :-)
Also, you'll have to sign a contributor's agreement at https://www.python.org/psf/contrib/contrib-form/

Thanks!

[mboquien]

I have now signed the contributor's agreement.

As for the unit test I was looking at it. However, I was wondering how to write a test that would have triggered the problem. It only shows up for very large arrays and it depends on occupied memory and the configuration of the temp dir. Or should I simply write a test creating for instance a 100 MB array and checking it has the right length?

[neologix]

Zero-filling mmap's backing file isn't really optimal: why not use truncate() instead? This way, it'll avoid completely I/O on filesystems that support sparse files, and should still work on FS that don't.

[mboquien]

If I remember correctly the problem is that some OS like linux (and probably others) do not really allocate space until something is written. If that's the case then the process may get killed later on when it writes something in the array.

Here is a quick example:

$ truncate -s 1T test.file
$ ls -lh test.file 
-rw-r--r-- 1 mederic users 1.0T Apr  2 23:10 test.file
$ df -h
Filesystem      Size  Used Avail Use% Mounted on
/dev/sdb1       110G   46G   59G  44% /home

[sbt]

Using truncate() to zero extend is not really portable: it is only guaranteed on XSI-compliant POSIX systems.

Also, the FreeBSD man page for mmap() has the following warning:

WARNING! Extending a file with ftruncate(2), thus creating a big
hole, and then filling the hole by modifying a shared mmap() can
lead to severe file fragmentation. In order to avoid such
fragmentation you should always pre-allocate the file's backing
store by write()ing zero's into the newly extended area prior to
modifying the area via your mmap(). The fragmentation problem is
especially sensitive to MAP_NOSYNC pages, because pages may be
flushed to disk in a totally random order.

[neologix]

If I remember correctly the problem is that some OS like linux (and
probably others) do not really allocate space until something is written.
If that's the case then the process may get killed later on when it writes
something in the array.

Yes, it's called overcommitting, and it's a good thing. It's exactly the
same thing for memory: malloc() can return non-NULL, and the process will
get killed when first writing to the page in case of memory pressure.

[mboquien]

"the process will get killed when first writing to the page in case of memory pressure."

According to the documentation, the returned shared array is zeroed. https://docs.python.org/3.4/library/multiprocessing.html#module-multiprocessing.sharedctypes

In that case because the entire array is written at allocation, the process is expected to get killed if allocating more memory than available. Unless I am misunderstanding something, which is entirely possible.

[neologix]

Also, the FreeBSD man page for mmap() has the following warning:

That's mostly important for real file-backed mapping.
In our case, we don't want a file-backed mmap: we expect the mapping to fit
entirely in memory, so the writeback/read performance isn't that important
to us.

Using truncate() to zero extend is not really portable: it is only
guaranteed on XSI-compliant POSIX systems.

Now that's annoying.
How about trying file.truncate() within a try block, and if an error is
raised fallback to the zero-filling?

Doing a lot of IO for an object which is supposed to be used for shared
memory is sad.

Or maybe it's time to add an API to access shared memory from Python (since
that's really what we're trying to achieve here).

According to the documentation, the returned shared array is zeroed.
In that case because the entire array is written at allocation, the
process is expected to get killed
if allocating more memory than available. Unless I am misunderstanding
something, which is entirely
possible.

Having the memory zero-filed doesn't require a write at all: when you do an
anonymous memory mapping for let's say 1Gb, the kernel doesn't
pre-emptively zero-fill it, it would be way to slow: usually it just sets
up the process page table to make this area a COW of a single zero page:
upon read, you'll read zeros, and upon write, it'll duplicate it as needed.

The only reason the code currently zero-fills the file is to avoid the
portability issues detailed by Richard.

[pitrou]

Or maybe it's time to add an API to access shared memory from Python
(since
that's really what we're trying to achieve here).

That sounds like a good idea. Especially since we now have the memoryview type.

[mboquien]

Thanks for the explanations Charles-François. I guess the new API would not be before 3.5 at least. Is there still a chance to integrate my patch (or any other) to improve the situation for the 3.4 series though?