Experimental memory allocator with vendored mimalloc

[tiran]

EXPERIMENTAL This is an experimental pull request and a proof of concept. mimalloc integration is in a very early stage.

Vendor and hook up mimalloc. The PR only replaces raw memory allocator and Python's internal obmalloc with mimalloc. It does not use any of the advanced features of mimalloc such as heaps.

• Includes mimalloc v1.7.1 and additional fixes
• --with-mimalloc=yes by default (for testing)
• --with-pymalloc=no by default
• --with-freelists=yes

mimalloc requires C11 <stdatomic.h> on POSIX platforms to provide atomic operations such as CAS. GCC on RHEL 7 / CentOS 7 is too old. GCC 4.9 and clang 3.6 seem to be minimum versions. Users need install Red Hat Developer Toolset 10.0 from Software Collections. The CentOS 7-based manylinux2014 container image comes with GCC 10.2 pre-installed.

related mimalloc tickets

• Fix strict function prototype warnings microsoft/mimalloc#467
• Define _DEFAULT_SOURCE for syscall and realpath microsoft/mimalloc#468
• Fix ARM64 MSVC linker problem microsoft/mimalloc#472
• Rename _os_random_weak to _mi_os_random_weak microsoft/mimalloc#474
• Add option to set mimalloc symbols to visibility=hidden microsoft/mimalloc#475

Initial benchmarks

Initial benchmarks should be taken with a big grain of salt. It seems like mimalloc is in the same ballpark as libc malloc + obmalloc with slightly better performance in some cases. I haven't measured memory consumption yet.

with mimalloc, without pymalloc, without freelists

In general about same performance, with some benchmarks 7% faster to 12% slower, https://gist.github.com/pablogsal/33fce1cbe5f908c1d85cae5d678f50a6

with mimalloc, without pymalloc, with freelists

In general 2% speedup, with some benchmarks 9% faster to 6% slower, https://gist.github.com/pablogsal/8027937b71cd30f17aaaa5ef7c885d3e

debug output

>>> sys._debugmallocstats()
mimalloc (version: 171)
heap stats:     peak      total      freed    current       unit      count  
  reserved:   256.0 mb   256.0 mb       0 b    256.0 mb       1 b              not all freed!
 committed:    20.0 mb    20.0 mb       0 b     20.0 mb       1 b              not all freed!
     reset:     1.0 mb     1.0 mb       0 b      1.0 mb       1 b              not all freed!
   touched:       0 b        0 b    932.3 kb  -932.3 kb       1 b              ok
  segments:       5          5          1          4                           not all freed!
-abandoned:       0          0          0          0                           ok
   -cached:       0          0          0          0                           ok
     pages:      94        123         29         94                           not all freed!
-abandoned:       0          0          0          0                           ok
 -extended:       0   
 -noretire:       0   
     mmaps:       0   
   commits:      69   
   threads:       0          0          0          0                           ok
  searches:     0.0 avg
numa nodes:       1
   elapsed:       5.572 s
   process: user: 0.073 s, system: 0.007 s, faults: 0, rss: 11.9 mb, commit: 20.0 mb
...

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @tiran for commit a4e71652eaacf491eb682d99fdbb0a521b8cb33f 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @tiran for commit b6f5965e657df821b2345433163293d5345d5bab 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @tiran for commit 11efb79362b5f49e441d3a24da01f3740a2222d5 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @tiran for commit 8d7373a0c3f1f47b1320087f8deafc7bd7c9ad4a 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[insilications]

Saw your PR in microsoft/mimalloc#475. This is great. I'm using mimalloc as a drop-in replacement with the ENV variables:

PYTHONMALLOC=malloc MIMALLOC_PAGE_RESET=0 MIMALLOC_LARGE_OS_PAGES=1 LD_PRELOAD='/usr/lib64/libmimalloc.so'

I expected that this upstream integration would bring other benefits, but it would not be able to use mimalloc's allocation of large OS pages?

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @tiran for commit 8d7373a0c3f1f47b1320087f8deafc7bd7c9ad4a 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[tiran]

This PR has been replaced by #31164.