The issue #23601 showed speedup for the dict type by replacing PyMem_Malloc() with PyObject_Malloc() in dictobject.c.

When I worked on the PEP 445, it was discussed to use the Python fast memory allocator for small memory allocations (<= 512 bytes), but I think that nobody tested on benchmark.

So I open an issue to discuss that.

By the way, we should also benchmark the Windows memory allocator which limits fragmentations. Maybe we can skip the Python small memory allocator on recent version of Windows?

Attached patch implements the change. The main question is the speedup on various kinds of memory allocations (need a benchmark) :-)

I will try to run benchmarks.

--

If the patch slows down Python, maybe we can investigate if some Python types (like dict) mostly uses "small" memory blocks (<= 512 bytes).

Ok, to avoid confusion, I opened an issue specific to Windows for its "Low-fragmentation Heap": issue #26251.

Other issues related to memory allocators.

Merged:

- issue #21233: Add *Calloc functions to CPython memory allocation API (extension of the PEP 445, asked by numpy)
- issue #13483: Use VirtualAlloc to allocate memory arenas (implementation of the PEP 445)
- issue #3329: API for setting the memory allocator used by Python

Open:

- issue #18835: Add aligned memory variants to the suite of PyMem functions/macros => this one is still open, the status is unclear :-/

Hum, the point of PyMem_Malloc() is that it's distinct from PyObject_Malloc(), right? Why would you redirect one to the other?

(of course, we might question why we have two different families of allocation APIs...)

> Hum, the point of PyMem_Malloc() is that it's distinct from PyObject_Malloc(), right? Why would you redirect one to the other?

For performances.

> (of course, we might question why we have two different families of allocation APIs...)

That's the real question: why does Python have PyMem family? Is it still justified in 2016?

--

Firefox uses jemalloc to limit the fragmentation of the heap memory. Once I spent a lot of time to try to understand the principle of fragmentation, and in my tiny benchmarks, jemalloc was *much* better than system allocator. By the way, jemalloc scales well on multiple threads ;-)

* http://www.canonware.com/jemalloc/
* https://github.com/jemalloc/jemalloc/wiki

My notes on heap memory fragmentation: http://haypo-notes.readthedocs.org/heap_fragmentation.html

About heap memory fragmentation, see also my attached two "benchmarks" in Python and C: python_memleak.py and tu_malloc.c.

So, I ran ssh://hg@hg.python.org/benchmarks with my patch. It looks like some benchmarks are up to 4% faster:

$ python3 -u perf.py ../default/python.orig ../default/python.pymem

INFO:root:Automatically selected timer: perf_counter
[ 1/10] 2to3...
INFO:root:Running `../default/python.pymem lib3/2to3/2to3 -f all lib/2to3`
INFO:root:Running `../default/python.pymem lib3/2to3/2to3 -f all lib/2to3` 1 time
INFO:root:Running `../default/python.orig lib3/2to3/2to3 -f all lib/2to3`
INFO:root:Running `../default/python.orig lib3/2to3/2to3 -f all lib/2to3` 1 time
[ 2/10] chameleon_v2...
INFO:root:Running `../default/python.pymem performance/bm_chameleon_v2.py -n 50 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_chameleon_v2.py -n 50 --timer perf_counter`
[ 3/10] django_v3...
INFO:root:Running `../default/python.pymem performance/bm_django_v3.py -n 50 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_django_v3.py -n 50 --timer perf_counter`
[ 4/10] fastpickle...
INFO:root:Running `../default/python.pymem performance/bm_pickle.py -n 50 --timer perf_counter --use_cpickle pickle`
INFO:root:Running `../default/python.orig performance/bm_pickle.py -n 50 --timer perf_counter --use_cpickle pickle`
[ 5/10] fastunpickle...
INFO:root:Running `../default/python.pymem performance/bm_pickle.py -n 50 --timer perf_counter --use_cpickle unpickle`
INFO:root:Running `../default/python.orig performance/bm_pickle.py -n 50 --timer perf_counter --use_cpickle unpickle`
[ 6/10] json_dump_v2...
INFO:root:Running `../default/python.pymem performance/bm_json_v2.py -n 50 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_json_v2.py -n 50 --timer perf_counter`
[ 7/10] json_load...
INFO:root:Running `../default/python.pymem performance/bm_json.py -n 50 --timer perf_counter json_load`
INFO:root:Running `../default/python.orig performance/bm_json.py -n 50 --timer perf_counter json_load`
[ 8/10] nbody...
INFO:root:Running `../default/python.pymem performance/bm_nbody.py -n 50 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_nbody.py -n 50 --timer perf_counter`
[ 9/10] regex_v8...
INFO:root:Running `../default/python.pymem performance/bm_regex_v8.py -n 50 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_regex_v8.py -n 50 --timer perf_counter`
[10/10] tornado_http...
INFO:root:Running `../default/python.pymem performance/bm_tornado_http.py -n 100 --timer perf_counter`
INFO:root:Running `../default/python.orig performance/bm_tornado_http.py -n 100 --timer perf_counter`

Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
6.880090 -> 6.818911: 1.01x faster

### fastpickle ###
Min: 0.453826 -> 0.442081: 1.03x faster
Avg: 0.456499 -> 0.443978: 1.03x faster
Significant (t=20.03)
Stddev: 0.00370 -> 0.00242: 1.5293x smaller

### fastunpickle ###
Min: 0.547908 -> 0.526027: 1.04x faster
Avg: 0.554663 -> 0.528686: 1.05x faster
Significant (t=15.95)
Stddev: 0.00893 -> 0.00728: 1.2260x smaller

### json_dump_v2 ###
Min: 2.733907 -> 2.627718: 1.04x faster
Avg: 2.762473 -> 2.664675: 1.04x faster
Significant (t=11.99)
Stddev: 0.03796 -> 0.04341: 1.1435x larger

### regex_v8 ###
Min: 0.042438 -> 0.042581: 1.00x slower
Avg: 0.042805 -> 0.044078: 1.03x slower
Significant (t=-2.12)
Stddev: 0.00171 -> 0.00388: 2.2694x larger

### tornado_http ###
Min: 0.254089 -> 0.246088: 1.03x faster
Avg: 0.257046 -> 0.249033: 1.03x faster
Significant (t=15.83)
Stddev: 0.00401 -> 0.00310: 1.2930x smaller

The following not significant results are hidden, use -v to show them:
chameleon_v2, django_v3, json_load, nbody.

real	19m13.413s
user	18m50.024s
sys	0m22.507s

> On Feb 2, 2016, at 7:00 AM, STINNER Victor <report@bugs.python.org> wrote:
> 
> So, I ran ssh://hg@hg.python.org/benchmarks with my patch. It looks like some benchmarks are up to 4% faster:

Please use -r flag for perf.py

> It looks like some benchmarks are up to 4% faster:

What this says is that some internals uses of PyMem_XXX should be replaced with PyObject_XXX.

FYI benchmark result to compare Python with and without pymalloc (fast memory allocator for block <= 512 bytes). As expected, no pymalloc is slower, up to 30% slower (and it's never faster).

Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
7.253671 -> 7.558993: 1.04x slower

### chameleon_v2 ###
Min: 5.598481 -> 5.794526: 1.04x slower
Avg: 5.714233 -> 5.922142: 1.04x slower
Significant (t=-8.01)
Stddev: 0.15956 -> 0.09048: 1.7636x smaller

### django_v3 ###
Min: 0.574221 -> 0.606462: 1.06x slower
Avg: 0.579659 -> 0.612088: 1.06x slower
Significant (t=-28.44)
Stddev: 0.00605 -> 0.00532: 1.1371x smaller

### fastpickle ###
Min: 0.450852 -> 0.502645: 1.11x slower
Avg: 0.455619 -> 0.513777: 1.13x slower
Significant (t=-26.24)
Stddev: 0.00696 -> 0.01404: 2.0189x larger

### fastunpickle ###
Min: 0.544064 -> 0.696306: 1.28x slower
Avg: 0.552459 -> 0.705372: 1.28x slower
Significant (t=-85.52)
Stddev: 0.00798 -> 0.00980: 1.2281x larger

### json_dump_v2 ###
Min: 2.780312 -> 3.265531: 1.17x slower
Avg: 2.830463 -> 3.370060: 1.19x slower
Significant (t=-23.73)
Stddev: 0.04190 -> 0.15521: 3.7046x larger

### json_load ###
Min: 0.428893 -> 0.558956: 1.30x slower
Avg: 0.431941 -> 0.569441: 1.32x slower
Significant (t=-74.76)
Stddev: 0.00791 -> 0.01033: 1.3060x larger

### regex_v8 ###
Min: 0.043439 -> 0.044614: 1.03x slower
Avg: 0.044388 -> 0.046487: 1.05x slower
Significant (t=-4.95)
Stddev: 0.00215 -> 0.00209: 1.0283x smaller

### tornado_http ###
Min: 0.264603 -> 0.278840: 1.05x slower
Avg: 0.270153 -> 0.285263: 1.06x slower
Significant (t=-23.04)
Stddev: 0.00489 -> 0.00436: 1.1216x smaller

The following not significant results are hidden, use -v to show them:
nbody.

Test with jemalloc using the shell script "python.jemalloc":
---
#!/bin/sh
LD_PRELOAD=/usr/lib64/libjemalloc.so /home/haypo/prog/python/default/python "$@"
---

Memory consumption:
python3 -u perf.py -m ../default/python ../default/python.jemalloc

Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
Mem max: 43100.000 -> 220.000: 195.9091x smaller

### chameleon_v2 ###
Mem max: 367276.000 -> 224.000: 1639.6250x smaller

### django_v3 ###
Mem max: 24136.000 -> 284.000: 84.9859x smaller

### fastpickle ###
Mem max: 8692.000 -> 284.000: 30.6056x smaller

### fastunpickle ###
Mem max: 8704.000 -> 216.000: 40.2963x smaller

### json_dump_v2 ###
Mem max: 10448.000 -> 216.000: 48.3704x smaller

### json_load ###
Mem max: 8444.000 -> 220.000: 38.3818x smaller

### nbody ###
Mem max: 7388.000 -> 220.000: 33.5818x smaller

### regex_v8 ###
Mem max: 12764.000 -> 220.000: 58.0182x smaller

### tornado_http ###
Mem max: 28216.000 -> 228.000: 123.7544x smaller



****

Performance:

Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
7.413484 -> 7.189792: 1.03x faster

### chameleon_v2 ###
Min: 5.559697 -> 5.869468: 1.06x slower
Avg: 5.672448 -> 6.033152: 1.06x slower
Significant (t=-13.67)
Stddev: 0.12098 -> 0.14203: 1.1740x larger

### nbody ###
Min: 0.242194 -> 0.229747: 1.05x faster
Avg: 0.244991 -> 0.235297: 1.04x faster
Significant (t=9.75)
Stddev: 0.00262 -> 0.00652: 2.4861x larger

### regex_v8 ###
Min: 0.042532 -> 0.046920: 1.10x slower
Avg: 0.043249 -> 0.047907: 1.11x slower
Significant (t=-13.23)
Stddev: 0.00180 -> 0.00172: 1.0503x smaller

### tornado_http ###
Min: 0.265755 -> 0.274526: 1.03x slower
Avg: 0.273617 -> 0.284186: 1.04x slower
Significant (t=-6.67)
Stddev: 0.00583 -> 0.01474: 2.5297x larger

The following not significant results are hidden, use -v to show them:
django_v3, fastpickle, fastunpickle, json_dump_v2, json_load.

>> It looks like some benchmarks are up to 4% faster:

> What this says is that some internals uses of PyMem_XXX should be replaced with PyObject_XXX.

Why not changing PyMem_XXX to use the same fast allocator than PyObject_XXX? (as proposed in this issue)

FYI we now also have the PyMem_RawXXX family :)

Le 02/02/2016 15:47, STINNER Victor a écrit :
> 
> ### 2to3 ###
> Mem max: 43100.000 -> 220.000: 195.9091x smaller
> 
> ### chameleon_v2 ###
> Mem max: 367276.000 -> 224.000: 1639.6250x smaller
> 
> ### django_v3 ###
> Mem max: 24136.000 -> 284.000: 84.9859x smaller

These figures are not even remotely believable.
It would make sense to investigate them before posting such numbers ;-)

Le 02/02/2016 15:48, STINNER Victor a écrit :
>> What this says is that some internals uses of PyMem_XXX should be replaced with PyObject_XXX.
> 
> Why not changing PyMem_XXX to use the same fast allocator than
PyObject_XXX? (as proposed in this issue)

Why have two sets of functions doing exactly the same thing?

> These figures are not even remotely believable.

To be honest, I didn't try to understand them :-) Are they the number of kB of the RSS memory?

Maybe perf.py doesn't like my shell script?

> Why have two sets of functions doing exactly the same thing?

I have no idea.

> Test with jemalloc using the shell script "python.jemalloc":
> ---
> #!/bin/sh
> LD_PRELOAD=/usr/lib64/libjemalloc.so /home/haypo/prog/python/default/python "$@"
> ---

"perf.py -m" doesn't work with such bash script, but it works using exec:
---
#!/bin/sh
LD_PRELOAD=/usr/lib64/libjemalloc.so exec /home/haypo/prog/python/default/python "$@"
---

> Memory consumption:
python3 -u perf.py -m ../default/python ../default/python.jemalloc


Hum, it looks like jemalloc uses *more* memory than libc memory allocators. I don't know if it's a known 


Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
Mem max: 43088.000 -> 43776.000: 1.0160x larger

### chameleon_v2 ###
Mem max: 367028.000 -> 626324.000: 1.7065x larger

### django_v3 ###
Mem max: 23824.000 -> 25120.000: 1.0544x larger

### fastpickle ###
Mem max: 8696.000 -> 9712.000: 1.1168x larger

### fastunpickle ###
Mem max: 8708.000 -> 9696.000: 1.1135x larger

### json_dump_v2 ###
Mem max: 10488.000 -> 11556.000: 1.1018x larger

### json_load ###
Mem max: 8444.000 -> 9396.000: 1.1127x larger

### nbody ###
Mem max: 7392.000 -> 8416.000: 1.1385x larger

### regex_v8 ###
Mem max: 12760.000 -> 13576.000: 1.0639x larger

### tornado_http ###
Mem max: 28196.000 -> 29920.000: 1.0611x larger

(Crap. I sent an incomplete message, sorry about that.)

> Hum, it looks like jemalloc uses *more* memory than libc memory allocators. I don't know if it's a known 

I don't know if it's a known issue/property of jemalloc.

Yury: "Please use -r flag for perf.py"

Oh, I didn't know this flag. Sure, I can do that.

New benchmark using --rigorous to measure the performance of attached pymem.patch.

It always seems faster, newer slower.

Report on Linux smithers 4.3.3-300.fc23.x86_64 #1 SMP Tue Jan 5 23:31:01 UTC 2016 x86_64 x86_64
Total CPU cores: 8

### 2to3 ###
Min: 6.772531 -> 6.686245: 1.01x faster
Avg: 6.875264 -> 6.726859: 1.02x faster
Significant (t=3.44)
Stddev: 0.09026 -> 0.03398: 2.6560x smaller

### django_v3 ###
Min: 0.562797 -> 0.552539: 1.02x faster
Avg: 0.591345 -> 0.557561: 1.06x faster
Significant (t=4.17)
Stddev: 0.07689 -> 0.02581: 2.9794x smaller

### fastpickle ###
Min: 0.464270 -> 0.437667: 1.06x faster
Avg: 0.467195 -> 0.442298: 1.06x faster
Significant (t=10.59)
Stddev: 0.01156 -> 0.02046: 1.7693x larger

### fastunpickle ###
Min: 0.548834 -> 0.526554: 1.04x faster
Avg: 0.554601 -> 0.539456: 1.03x faster
Significant (t=4.67)
Stddev: 0.01137 -> 0.03040: 2.6734x larger

### json_dump_v2 ###
Min: 2.723152 -> 2.603108: 1.05x faster
Avg: 2.749255 -> 2.693655: 1.02x faster
Significant (t=2.89)
Stddev: 0.03016 -> 0.18988: 6.2963x larger

### regex_v8 ###
Min: 0.044256 -> 0.042201: 1.05x faster
Avg: 0.044733 -> 0.043134: 1.04x faster
Significant (t=4.55)
Stddev: 0.00201 -> 0.00288: 1.4309x larger

### tornado_http ###
Min: 0.253405 -> 0.247401: 1.02x faster
Avg: 0.256274 -> 0.250380: 1.02x faster
Significant (t=17.48)
Stddev: 0.00285 -> 0.00382: 1.3430x larger

The following not significant results are hidden, use -v to show them:
chameleon_v2, json_load, nbody.

Hi all,

Please find below the results from a complete GUPB run on a patched CPython 3.6. In average, an improvement of about 2.1% can be observed. 

I'm also attaching an implementation of the patch for CPython 2.7 and its benchmark results. On GUPB the average performance boost is 1.5%. 
In addition we are also seeing a 2.1% increase in throughput rate from our OpenStack Swift setup as measured by ssbench.

Compared to my proposition in issue #26382, this patch yields slightly better results for CPython 3.6, gaining an average of +0.36% on GUPB,
and similar results for CPython 2.7.


Hardware and OS configuration:
==============================
Hardware:           Intel XEON (Haswell-EP)

BIOS settings:      Intel Turbo Boost Technology: false
                    Hyper-Threading: false                  

OS:                 Ubuntu 14.04.2 LTS

OS configuration:   Address Space Layout Randomization (ASLR) disabled to reduce run
                    to run variation by echo 0 > /proc/sys/kernel/randomize_va_space
                    CPU frequency set fixed at 2.3GHz

Repository info:
================
CPython2 : 2d8e8d0e7162 (2.7)
CPython3 : f9391e2b74a5 tip

Results
=======

Table 1: CPython 3 GUPB results
-------------------------------
unpickle_list           22.74%
mako_v2                 9.13%
nqueens                 6.32%
meteor_contest          5.61%
fannkuch                5.34%
simple_logging          5.28%
formatted_logging       5.06%
fastunpickle            4.37%
json_dump_v2            3.10%
regex_compile           3.01%
raytrace                2.95%
pathlib                 2.43%
tornado_http            2.22%
django_v3               1.94%
telco                   1.65%
pickle_list             1.59%
chaos                   1.50%
etree_process           1.48%
fastpickle              1.34%
silent_logging          1.12%
2to3                    1.09%
float                   1.01%
nbody                   0.89%
normal_startup          0.86%
startup_nosite          0.79%
richards                0.67%
regex_v8                0.61%
etree_generate          0.57%
hexiom2                 0.54%
pickle_dict             0.20%
call_simple             0.18%
spectral_norm           0.17%
regex_effbot            0.16%
unpack_sequence         0.00%
call_method_unknown    -0.04%
chameleon_v2           -0.07%
json_load              -0.08%
etree_parse            -0.09%
pidigits               -0.15%
go                     -0.16%
etree_iterparse        -0.22%
call_method_slots      -0.49%
call_method            -0.97%


Table 2: CPython 2 GUPB results
-------------------------------
unpickle_list           16.88%
json_load               11.74%
fannkuch                8.11%
mako_v2                 6.91%
meteor_contest          6.27%
slowpickle              4.81%
nqueens                 4.46%
html5lib_warmup         3.53%
chaos                   2.67%
regex_v8                2.56%
html5lib                2.34%
fastunpickle            2.32%
tornado_http            2.23%
rietveld                2.15%
simple_logging          1.82%
normal_startup          1.57%
call_method_slots       1.53%
telco                   1.49%
regex_compile           1.47%
spectral_norm           1.36%
hg_startup              1.27%
regex_effbot            1.18%
nbody                   1.02%
2to3                    1.01%
pybench                 0.99%
chameleon_v2            0.98%
slowunpickle            0.93%
startup_nosite          0.92%
pickle_list             0.89%
richards                0.56%
django_v3               0.48%
json_dump_v2            0.41%
raytrace                0.38%
unpack_sequence         0.00%
float                  -0.05%
slowspitfire           -0.07%
go                     -0.24%
hexiom2                -0.26%
spambayes              -0.27%
pickle_dict            -0.30%
etree_parse            -0.32%
pidigits               -0.41%
etree_iterparse        -0.47%
bzr_startup            -0.55%
fastpickle             -0.74%
etree_process          -0.96%
formatted_logging      -1.01%
call_simple            -1.08%
pathlib                -1.12%
silent_logging         -1.22%
etree_generate         -1.23%
call_method_unknown    -2.14%
call_method            -2.22%

Table 3: OpenStack Swift ssbench results
----------------------------------------
ssbench                 2.11%

> Compared to my proposition in issue #26382, this patch yields slightly better results for CPython 3.6, gaining an average of +0.36% on GUPB,
and similar results for CPython 2.7.

IMHO this change is too young to be backported to Python 2.7. I wrote it for Python 3.6 only. For Python 2.7, I suggest to write patches with narrow scope, as you did for the patch only modifying the list type.

"""
Table 1: CPython 3 GUPB results
-------------------------------
unpickle_list           22.74%
mako_v2                 9.13%
nqueens                 6.32%
meteor_contest          5.61%
fannkuch                5.34%
simple_logging          5.28%
formatted_logging       5.06%
"""

I surprised to see slow-down, but I prefer to think that changes smaller than 5% are pure noise.

The good news is the long list of benchmarks with speedup larger than 5.0% :-) 22% on unpick list is nice to have too!

I've just posted the results to an OpenStack Swift benchmark run using the patch from my proposition, issue #26382. 
Victor's patch, applied to CPython 2.7, adds an extra 1% compared to mine (which improved throughput by 1%), effectively doubling the performance gain. Swift is a highly complex real-world workload, so this result is quite significant.

I created the issue #26516 "Add PYTHONMALLOC env var and add support for malloc debug hooks in release mode" to help developers to detect bugs in their code, especially misuse of the PyMem_Malloc() API.

Patch 3:

- Ooops, I updated pymem_api_misuse(), but I forgot to update the related unit test. It's now fixed.

In february 2016, I started a thread on the python-dev mailing list:
[Python-Dev] Modify PyMem_Malloc to use pymalloc for performance
https://mail.python.org/pipermail/python-dev/2016-February/143084.html

M.-A. Lemburg wrote:

"""

> Do you see any drawback of using pymalloc for PyMem_Malloc()?

Yes: You cannot free memory allocated using pymalloc with the
standard C lib free().

It would be better to go through the list of PyMem_*() calls
in Python and replace them with PyObject_*() calls, where
possible.

> Does anyone recall the rationale to have two families to memory allocators?

The PyMem_*() APIs were needed to have a cross-platform malloc()
implementation which returns standard C lib free()able memory,
but also behaves well when passing 0 as size.
"""


M.-A. Lemburg fears that the PyMem_Malloc() API is misused:

"""

Sometimes, yes, but we also do allocations for e.g.
parsing values in Python argument tuples (e.g. using
"es" or "et"):

https://docs.python.org/3.6/c-api/arg.html

We do document to use PyMem_Free() on those; not sure whether
everyone does this though.
"""


M.-A. Lemburg suggested to the patch of this issue on:

"""
Yes, but those are part of the stdlib. You'd need to check
a few C extensions which are not tested as part of the stdlib,
e.g. numpy, scipy, lxml, pillow, etc. (esp. ones which implement custom
types in C since these will often need the memory management
APIs).

It may also be a good idea to check wrapper generators such
as cython, swig, cffi, etc.
"""

numpy: good!

* I patched Pyhon 3.6 with pymem.patch of this issue + pymem-3.patch of issue #26516
* I had issues to run tests with Python 3.6 compiled in debug mode: http://bugs.python.org/issue26519 & https://github.com/numpy/numpy/issues/7399
* I ran the test suite: all tests pass, no bug related to memory allocators
* Tested numpy version: commit b92cc76afad2e74cbbf6f5b9f5b68050f7c8642a (Mar 7 2016)

Commands ran in numpy tests in a virtual environment:

numpy$ python setup.py install
numpy$ cd..
$ python -c 'import numpy; numpy.test()'
(...)
Ran 6206 tests in 280.986s

OK (KNOWNFAIL=7, SKIP=6)

Victor, why do you insist on this instead of changing internal API calls in CPython?

Antoine Pitrou added the comment:
> Victor, why do you insist on this instead of changing internal API calls in CPython?

https://mail.python.org/pipermail/python-dev/2016-February/143097.html

"There are 536 calls to the functions PyMem_Malloc(), PyMem_Realloc()
and PyMem_Free()."

> "There are 536 calls to the functions PyMem_Malloc(), PyMem_Realloc()
and PyMem_Free()."

I'm sure you can use powerful tools such as "sed" ;-)

> I'm sure you can use powerful tools such as "sed" ;-)

I guess that PyMem functions are used in third party C extensions modules. I expect (minor) speedup in these modules too.

I don't understand why we should keep a slow allocator if Python has a faster allocator?

lxml: good!

* I patched Python 3.6 with pymem.patch of this issue + pymem-3.patch of issue #26516
* Tested lxml version: git commit 93ec66f6533995a7742278f9ba14b925149ac140 (Mar 8 2016)

lxml$ make test
(...)
Ran 1735 tests in 27.663s

OK

Pillow: good

Note: I had to install JPEG headers (sudo dnf install -y libjpeg-turbo-devel).

Tested version: git commit 555544c5cfc3874deaac9cfa87780822ee714c0d (Mar 8 2016).

---
Pillow$ python setup.py install
Pillow$ python selftest.py
Pillow$ python  test-installed.py
(...)
Ran 671 tests in 8.458s

FAILED (SKIP=124, errors=2)
---

The two errors are "OSError: decoder libtiff not available".

Le 09/03/2016 18:01, STINNER Victor a écrit :
> I don't understand why we should keep a slow allocator if Python has a faster allocator?

Define "slow".  malloc() on Linux should be reasonably fast.

Do you think it's reasonable to risk breaking external libraries just
for a hypothetic "performance improvement"?

Again, why don't you try simply changing internal calls?

> Define "slow".  malloc() on Linux should be reasonably fast.

See first messages of this issue for benchmark results. Some specific benchmarks are faster, none is slower.


> Do you think it's reasonable to risk breaking external libraries just
for a hypothetic "performance improvement"?

Yes. It was discussed in the python-dev thread.

> Yes. It was discussed in the python-dev thread.

I'm talking about the performance improvement in third-party libraries, not the performance improvement in CPython itself which can be addressed by replacing the internal API calls.

> I'm talking about the performance improvement in third-party libraries, not the performance improvement in CPython itself which can be addressed by replacing the internal API calls.

Oh ok. I don't know how to measure the performance of third-party libraries. I expect no speedup or a little speedup, but no slow-down.


> Do you think it's reasonable to risk breaking external libraries just
for a hypothetic "performance improvement"?

The question is if my change really breaks anything in practice. I'm testing some popular C extensions to prepare an answer. Early results is that developer use correctly the Python allocator API :-)

I disagree on the fact that my change breaks any API. The API doc is clear. For example, you must use PyMem_Free() on memory allocated by PyMem_Malloc(). If you use free(), it fails badly with Python compiled in debug mode.

My issue #26516 "Add PYTHONMALLOC env var and add support for malloc debug hooks in release mode" may help developers to validate their own application.

I suggest you to continue the discussion on python-dev for a wider audience. I will test a few more projects before replying on the python-dev thread.

Le 09/03/2016 18:27, STINNER Victor a écrit :
> 
> I disagree on the fact that my change breaks any API. The API doc is
clear.

Does the API doc say anything about the GIL, for example? Or Valgrind?

> I suggest you to continue the discussion on python-dev for a wider
audience. I will test a few more projects before replying on the
python-dev thread.

I have no interest in going back and forth between the Python tracker
and python-dev (especially since I hardly read python-dev these days).
If you address my questions positively here I will be happy with the patch!

cryptography: good

* Git commit 0681de7241dcbaec7b3dc85d3cf3944e4bec8309 (Mar 9 2016)

"4 failed, 77064 passed, 3096 skipped in 405.09 seconds" 

1 error is related to the version number (probably an issue on how I run the tests), 3 errors are FileNotFoundError related to cryptography_vectors. At least, there is no Python fatal error related to memory allocators ;-)

--

Hum, just in case, I checked my venv:

(ENV) haypo@smithers$ python -c 'import _testcapi; _testcapi.pymem_api_misuse()'
...
Fatal Python error: bad ID: Allocated using API 'o', verified using API 'r'

(ENV) haypo@smithers$ python -c 'import _testcapi; _testcapi.pymem_buffer_overflow()'
...
Fatal Python error: bad trailing pad byte

It works ;-)

2016-03-09 18:28 GMT+01:00 Antoine Pitrou <report@bugs.python.org>:
> Does the API doc say anything about the GIL, for example? Or Valgrind?

For the GIL, yes, Python 3 doc is explicit:
https://docs.python.org/dev/c-api/memory.html#memory-interface

Red and bold warning: "The GIL must be held when using these functions."

Hum, sadly it looks like the warning miss in Python 2 doc.

The GIL was the motivation to introduce the PyMem_RawMalloc() function
in Python 3.4.

For Valgrind: using the issue #26516, you will be able to use
PYTHONMALLOC=malloc to use easily Valgrind even on a Python compiled
in release mode (which is a new feature, before you had to manually
recompile Python in debug mode with --with-valgrind)).

> Does the API doc say anything about the GIL, for example?

I modified Python to add assert(PyGILState_Check()); in PyMem_Malloc() and other functions.

Sadly, I found a bug in Numpy: Numpy releases the GIL for performance but call PyMem_Malloc() with the GIL released. I proposed a fix:
https://github.com/numpy/numpy/pull/7404

I guess that the fix is obvious and will be quickly merged, but it means that other libraries may have the issue.

Using the issue #26516 (PYTHONMALLOC=debug), we can check PyGILState_Check() at runtime, but there is currently an issue related to sub-interpreters. The assertion fails in support.run_in_subinterp(), function used by test_threading and test_capi for example.

pymalloc.patch: Updated patch.

> Using the issue #26516 (PYTHONMALLOC=debug), we can check PyGILState_Check() at runtime, but there is currently an issue related to sub-interpreters. The assertion fails in support.run_in_subinterp(), function used by test_threading and test_capi for example.

I created #26558 to implement GIL checks in PyMem_Malloc() and PyObject_Malloc().

I created the issue #26563 "PyMem_Malloc(): check that the GIL is hold in debug hooks".

New changeset 68b2a43d8653 by Victor Stinner in branch 'default':
PyMem_Malloc() now uses the fast pymalloc allocator
https://hg.python.org/cpython/rev/68b2a43d8653

New changeset 104ed24ebbd0 by Victor Stinner in branch 'default':
Issue #26249: Try test_capi on Windows
https://hg.python.org/cpython/rev/104ed24ebbd0

New changeset 7acad5d8f80e by Victor Stinner in branch 'default':
Issue #26249: Mention PyMem_Malloc() change in What's New in Python 3.6 in the
https://hg.python.org/cpython/rev/7acad5d8f80e

I documented the change, buildbots are happy, I close the issue.

68b2a43d8653 introduced memory leak.

$ ./python -m test.regrtest -uall -R : test_format
Run tests sequentially
0:00:00 [1/1] test_format
beginning 9 repetitions
123456789
.........
test_format leaked [6, 7, 7, 7] memory blocks, sum=27
1 test failed:
    test_format
Total duration: 0:00:01

New changeset 090502a0c69c by Victor Stinner in branch 'default':
Issue #25349, #26249: Fix memleak in formatfloat()
https://hg.python.org/cpython/rev/090502a0c69c

> 68b2a43d8653 introduced memory leak.

I was very surprised to see a regression in test_format since I didn't change any change related to bytes, bytearray or str formatting in this issue.

In fact, it's much better than that! With PyMem_Malloc() using pymalloc, we benefit for free of the cheap "_Py_AllocatedBlocks" memory leak detector. I introduced the memory leak in the issue #25349 when I optimimzed bytes%args and bytearray%args using the new _PyBytesWriter API.

This memory leak gave me an idea, I opened the issue #26850: "PyMem_RawMalloc(): update also sys.getallocatedblocks() in debug mode".

There are no more know bugs related to this change, I close the issue. Thanks for the test_format report Serhiy, I missed it.