Re: [Python-Dev] mmap & munmap loop (Was: Compact ordered set

Maybe pickle is inefficient in its memory management and causes a lot of memory fragmentation? It's hard to write an efficient memory allocator :-( My notes on memory: * "Excessive peak memory consumption by the Python parser" https://bugs.python.org/issue26415 * https://pythondev.readthedocs.io/memory.html * https://vstinner.readthedocs.io/heap_fragmentation.html Sometimes I would like to be able to use a separated memory allocator for one function, to not pollute the global allocator and so avoid "punching holes" in global memory pools or in the heap memory. The problem is to track the lifetime of objects. If the allocated objects live longer than the function, PyObject_Free() should be able to find the alllocator used by the memory block. pymalloc is already able to check if its manages a memory block using its address. If it's not allocated by pymalloc, PyObject_Free() falls back to libc free(). The Python parser already uses PyArena which is a custom memory allocator. It uses PyMem_Malloc() to allocate memory. In Python 3.7, PyMem_Malloc() uses pymalloc: https://docs.python.org/dev/c-api/memory.html#default-memory-allocators Victor Le mer. 27 févr. 2019 à 12:36, INADA Naoki a écrit :

It happened very accidentally. Since venv is used, many paths in the interpreter is changed. So how memory is used are changed.

Let's reproduce the accident.

$ cat m2.py import pickle, sys

LIST = pickle.dumps([[0]*10 for _ in range(10)], pickle.HIGHEST_PROTOCOL)

N = 1000 z = [[0]*10 for _ in range(N)]

if '-c' in sys.argv: sys._debugmallocstats() sys.exit()

for _ in range(100000): pickle.loads(LIST)

$ /usr/bin/time python3 m2.py 0.42user 0.00system 0:00.43elapsed 99%CPU (0avgtext+0avgdata 9100maxresident)k 0inputs+0outputs (0major+1139minor)pagefaults 0swaps

There are only 1139 faults. It is less than 100000.

$ /usr/bin/time python3 m2.py -c ... 14 unused pools * 4096 bytes = 57,344 ...

adjust N im m2.py until it shows "0 unused pools". In my case, N=1390.

$ /usr/bin/time python3 m2.py 0.51user 0.33system 0:00.85elapsed 99%CPU (0avgtext+0avgdata 9140maxresident)k 0inputs+0outputs (0major+201149minor)pagefaults 0swaps

200000 faults! It seems two page fault / loop. (2 pools are used and returned).

On Wed, Feb 27, 2019 at 7:51 PM Victor Stinner wrote:

...

Sorry, I didn't get a coffee yet: more *often* in a venv.

Le mer. 27 févr. 2019 à 11:32, Victor Stinner a écrit :

...

Any idea why Python calls mmap+munmap more even in a venv?

Victor

Le mer. 27 févr. 2019 à 10:00, INADA Naoki a écrit :

...

...

...

Ah, another interesting point, this huge slowdown happens only when bm_pickle.py is executed through pyperformance. When run it directly, slowdown is not so large.

pyperformance runs benchmarks in a virtual environment. I don't know if it has any impact on bm_pickle.

Most pyperformance can be run outside a virtual env if required modules are installed on the system. (bm_pickle only requires the stdlib and perf.)

Bingo!

Without venv:

unpickle: Mean +- std dev: 26.9 us +- 0.0 us % time seconds usecs/call calls errors syscall ------ ----------- ----------- --------- --------- ---------------- 28.78 0.000438 0 1440 read 27.33 0.000416 1 440 25 stat 9.72 0.000148 1 144 mmap ... 0.79 0.000012 1 11 munmap

With venv:

% time seconds usecs/call calls errors syscall ------ ----------- ----------- --------- --------- ---------------- 57.12 0.099023 2 61471 munmap 41.87 0.072580 1 61618 mmap 0.23 0.000395 1 465 27 stat

unpickle and unpickle_list creates massive same-sized objects, then all objects are removed. If all pools in the arena is freed, munmap is called.

I think we should save some arenas to reuse. On recent Linux, we may be able to use MADV_FREE instead of munmap.

-- Night gathers, and now my watch begins. It shall not end until my death.

-- Night gathers, and now my watch begins. It shall not end until my death.

-- INADA Naoki

-- Night gathers, and now my watch begins. It shall not end until my death.