From https://github.com/microsoft/mimalloc

mimalloc (pronounced "me-malloc") is a general purpose allocator with excellent performance characteristics. Initially developed by Daan Leijen for the run-time systems of the Koka and Lean languages.

mimalloc has several interesting properties that make it useful for CPython. Amongst other it is fast, thread-safe, and NUMA-aware. It has built-in free lists with multi-sharding and allocation heaps. While Python's obmalloc requires the GIL to protect its data structures, mimalloc uses mostly thread-local and atomic instructions (compare-and-swap) for efficiency. Sam Gross' nogil relies on mimalloc's thread safety and uses first-class heaps for heap walking GC.

mimalloc works on majority of platforms and CPU architectures. However it requires a compiler with C11 atomics support. CentOS 7's default GCC is slightly too old, more recent GCC from Developer Toolset is required.

For 3.11 I plan to integrate mimalloc as an optional drop-in replacement for obmalloc. Users will be able to compile CPython without mimalloc or disable mimalloc with PYTHONMALLOC env var. Since mimalloc will be optional in 3.11, Python won't depend or expose on any of the advanced features yet. The approach enables the community to test and give feedback with minimal risk of breakage.

mimalloc sources will vendored without any option to use system libraries. Python's mimalloc requires several non-standard compile-time flags. In the future Python may extend or modify mimalloc for heap walking and nogil, too.

(This is a tracking bug until I find time to finish a PEP.)

I add Neil to the nosy list since he is one of the kick-off members with this amazing works :)

New features:

• vendored mimalloc 2.0.3 + two patches from mimalloc dev branch. Mimalloc is embedded in obmalloc.o. Symbols are either hidden or names are mangled to have a _Py_ prefix.
• ./configure --with[out]-mimalloc (default: yes), fails if atomics are not available.
• PYTHONMALLOC=mimalloc, PYTHONMALLOC=mimalloc-debug env var settings
• PYMEM_ALLOCATOR_MIMALLOC, PYMEM_ALLOCATOR_MIMALLOC_DEBUG
• sys.debugmallocstats() and _PyObject_DebugMallocStats() prints mimalloc stats
• sys._malloc_info struct, contains information about available and current allocator

Buildbots "PPC64 Fedora PR" and all RHEL 7 build bots provided by David Edelsohn are failing because compiler is missing support for stdatomic.h.

Thanks, I'm indeed interested. Most credit goes to Christian for advancing this.

For the missing stdatomic.h, would it be appropriate to have an autoconfig check for it? Can just disable mimalloc if it doesn't exist.

We have an autoconf check for stdatomic.h. The test even verifies that a program with atomic_load_explicit() compiles and links.

How do we want to use mimalloc in the future? Is it going to stay optional in 3.12? Then the default setting for --with-mimalloc should depend on presence of stdatomic.h. Do we want to make it mandatory for GC heap walking and nogil? Then --with-mimalloc should default to "yes" and configure should abort when stdatomic.h is missing.

I'm leaning towards --with-mimalloc=yes. It will make users aware that they need a compiler with atomics:

configure: error: --with-mimalloc requires stdatomic.h. Update your compiler or rebuild with --without-mimalloc. Python 3.12 will require stdatomic.

ICC might be a problem. Apparently some version have an incomplete stdatomic.h, see bpo-37415.

References:

• C11 stdatomic.h https://en.cppreference.com/w/c/atomic
• mimalloc-atomic.h https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-atomic.h
• MSVC Interlocked Variable Access https://docs.microsoft.com/de-de/windows/win32/sync/interlocked-variable-access

My preference would be for --with-mimalloc=yes in an upcoming release. For platforms without the required stdatomic.h stuff, they can manually specify --with-mimalloc=no. That will make them aware that a future release of Python might no longer build (if mimalloc is no longer optional).

A soft-landing for merging nogil is not a good enough reason to merge mimalloc, IMHO. nogil may never be merged. There should be some concrete and immediate advantage to switch to mimalloc. The idea of using the "heap walking" to improve is cyclic GC is not concrete enough. It's just an idea at this point.

I think the (small) performance win could be enough of a reason to merge. This seems to be the most recent benchmark:

https://gist.github.com/pablogsal/8027937b71cd30f17aaaa5ef7c885d3e

There is also the long-term maintenance issue. So far, mimalloc upstream has been responsive. The mimalloc code is not so huge or complicated that we couldn't maintain it (if for some reason it gets abandoned upstream). However, I think we would prefer to maintain obmalloc rather than mimalloc, all else being equal. Abandonment by the upstream seems fairly unlikely. So, I'm not too concerned about maintenance.

New benchmark:

Benchmark hidden because not significant (9): unpack_sequence, go, raytrace, chameleon, xml_etree_process, fannkuch, sqlite_synth, regex_effbot, unpickle_list

I re-ran the benchmark of d6f5f010b586:

Benchmark hidden because not significant (7): scimark_fft, dulwich_log, python_startup_no_site, regex_effbot, sqlite_synth, nbody, pickle_list

ICC 2021 has full support for stdatomic.h and compiles mimalloc just fine:

$ CC="icc" ./configure -C --with-pydebug
$ make
$ ./python
Python 3.11.0a5+ (main, Feb  9 2022, 15:57:40) [GCC Intel(R) C++ gcc 7.5 mode] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> sys._malloc_info
sys._malloc_info(allocator='mimalloc_debug', with_pymalloc=True, with_mimalloc=True, mimalloc_secure=4, mimalloc_debug=2)

AIX xlc is still a problem. It does not support C11 stdatomic.h. But it comes with older GCC atomic memory access __sync function family, https://www.ibm.com/docs/en/xl-c-and-cpp-aix/13.1.3?topic=cbif-gcc-atomic-memory-access-built-in-functions-extension . It might be possible to re-implement mimalloc's atomics with __sync functions (e.g. https://gist.github.com/nhatminhle/5181506). The implementation would be less efficient, though. The __sync functions don't have memory order, atomic_load_explicit(v) becomes __sync_fetch_and_add(v, 0), and atomic_store_explicit() requires two full memory barriers.

FYI, PR #31164 indicates the following 3 blockers:

• mimalloc is holding on to large blocks of memory (thanks @neonene)
• the ASAN CI job is failing
• PGO on Windows is failing

After a conversation between @tiran, @daanx (the mimalloc creator), and me, we have a good idea on how to move forward. I'm looking into the PGO issue (with some help).

[...]
There is also the long-term maintenance issue. So far, mimalloc upstream has been responsive. The mimalloc code is not so huge or complicated that we couldn't maintain it (if for some reason it gets abandoned upstream). However, I think we would prefer to maintain obmalloc rather than mimalloc, all else being equal. Abandonment by the upstream seems fairly unlikely. So, I'm not too concerned about maintenance.

I made a similar point wrt the bundling at #109914 (comment).

I think this issue could be closed now that mimalloc went into HEAD already.

[...]
There is also the long-term maintenance issue. So far, mimalloc upstream has been responsive. The mimalloc code is not so huge or complicated that we couldn't maintain it (if for some reason it gets abandoned upstream). However, I think we would prefer to maintain obmalloc rather than mimalloc, all else being equal. Abandonment by the upstream seems fairly unlikely. So, I'm not too concerned about maintenance.

I made a similar point wrt the bundling at #109914 (comment).

See follow-up issue:

• Mimalloc differences from upstream #113141