[corona10]

• Issue: Mimalloc differences from upstream #113141

[DinoV]

https://github.com/facebookexperimental/free-threading-benchmarking/actions/runs/39860993693

[corona10]

Based on an analysis of python -c "import re", with help from a script
written with Claude and run under gdb.

The effect of reducing the page size differs by object size.

Scope	AS-IS pages	AS-IS resident	AS-IS live	AS-IS util	TO-BE pages	TO-BE resident	TO-BE live	TO-BE util
All heaps	149	5480 KiB	4106 KiB	75 %	309	4964 KiB	4102 KiB	83 %
MEM	46	1668 KiB	1146 KiB	69 %	77	1420 KiB	1142 KiB	80 %
OBJECT	45	1604 KiB	1083 KiB	68 %	88	1344 KiB	1083 KiB	81 %
GC	47	2108 KiB	1841 KiB	87 %	133	2088 KiB	1841 KiB	88 %
GC_PRE	11	100 KiB	35 KiB	35 %	11	112 KiB	35 KiB	32 %
<=256 B	68	2540 KiB	2228 KiB	88 %	174	2556 KiB	2228 KiB	87 %
256 B-2 KiB	47	1636 KiB	1346 KiB	82 %	109	1608 KiB	1346 KiB	84 %
2-64 KiB	30	1088 KiB	315 KiB	29 %	22	584 KiB	311 KiB	53 %
>64 KiB	4	216 KiB	216 KiB	100 %	4	216 KiB	216 KiB	100 %

[corona10]

cc @nascheme

[corona10]

@DinoV @nascheme @mpage
According to the benchmark results, this change reduces memory usage by 3% compared to the baseline, with no performance regression. Although there is no comparison with the default build, the results do confirm a reduction in memory usage.

[corona10]

cc @daanx

[corona10]

We may need to tune a 32-bit system, but I prefer to pass at this time.

[corona10]

@nascheme @DinoV
The results from Tools/gcbench from Thomas show that reducing the page size from 64 KiB to 16 KiB increases GC time by 6–8%, since each collection visits four times as many pages. GC-bound pathological cases are 5–8% slower, while steady-state workloads and pyperformance show no change in total runtime.

Peak RSS is generally unchanged on these large-heap tests, although bigcollect/container shows an 11% increase. That case would be worth a separate follow-up if we decide to merge this PR.

This PR aims to reduce the memory overhead relative to the default build from the current 18% to 15%, as measured by pyperformance. I think we should also compare against the default build using the GC benchmarks to make sure both the performance and memory overheads remain within the guardrails outlined in PEP 779. If both criteria are met, I would like to propose merging this PR.

I’ll share the results once I’ve run the GC benchmarks against the default build.

[corona10]

Even in GC benchmark this tune does not harmful to the current free-threaded build
(I re-oreganize as pyperf style result)

Benchmarks with tag 'pathology': time

Benchmark	default	FT 64K	FT 16K
gen_frag sparse	583 ms	1.12 sec: 1.92x slower	1.22 sec: 2.10x slower
gen_frag churn	5.07 sec	3.60 sec: 1.41x faster	3.97 sec: 1.28x faster
gen_frag churn acyclic	4.63 sec	2.83 sec: 1.64x faster	2.95 sec: 1.57x faster
bigcollect	104 ms	97 ms: 1.07x faster	103 ms: 1.01x faster
container	446 ms	382 ms: 1.17x faster	399 ms: 1.12x faster
doctree	6.19 sec	4.67 sec: 1.33x faster	4.90 sec: 1.26x faster
migrations	6.62 sec	5.02 sec: 1.32x faster	5.21 sec: 1.27x faster
deque	555 ms	774 ms: 1.40x slower	751 ms: 1.35x slower
bigbuffers 200KB (†)	51 ms	103 ms: 2.00x slower	96 ms: 1.87x slower
bigbuffers 4MB (†)	3.34 sec	1.23 sec: 2.72x faster	417 ms: 8.01x faster
ext_buffers_malloc (*)	549,634 ops/s	337,796 ops/s: 1.63x slower	338,180 ops/s: 1.63x slower
ext_buffers_pymem	558,001 ops/s	325,207 ops/s: 1.72x slower	301,579 ops/s: 1.85x slower
hidden_refs	1,589,119 ops/s	470,105 ops/s: 3.38x slower	499,478 ops/s: 3.18x slower
c_globals	291,985 ops/s	227,908 ops/s: 1.28x slower	223,261 ops/s: 1.31x slower
py_globals	283,106 ops/s	241,875 ops/s: 1.17x slower	235,612 ops/s: 1.20x slower
event_widgets runaway (*)	33,171 ops/s	45,872 ops/s: 1.38x faster	41,341 ops/s: 1.25x faster
Geometric mean	(ref)	1.13x slower	1.17x slower
Geometric mean incl. (†)	(ref)	1.09x slower	1.05x slower

Benchmarks with tag 'pathology': peak memory

Benchmark	default	FT 64K	FT 16K
gen_frag sparse	49 MB	54 MB: 1.10x	54 MB: 1.11x
gen_frag churn	90 MB	88 MB: 0.98x	88 MB: 0.98x
gen_frag churn acyclic	88 MB	88 MB: 1.00x	88 MB: 1.00x
bigcollect	260 MB	275 MB: 1.06x	305 MB: 1.17x
container	245 MB	273 MB: 1.11x	304 MB: 1.24x
doctree	1,165 MB	1,208 MB: 1.04x	1,225 MB: 1.05x
migrations	277 MB	298 MB: 1.08x	301 MB: 1.09x
deque	246 MB	105 MB: 0.43x	91 MB: 0.37x
bigbuffers 200KB	204 MB	418 MB: 2.05x	420 MB: 2.06x
bigbuffers 4MB	1,907 MB	2,017 MB: 1.06x	2,031 MB: 1.07x
ext_buffers_malloc (*)	188 MB	392 MB: 2.08x	508 MB: 2.70x
ext_buffers_pymem	186 MB	439 MB: 2.36x	441 MB: 2.37x
hidden_refs	306 MB	420 MB: 1.37x	424 MB: 1.39x
c_globals	124 MB	170 MB: 1.37x	173 MB: 1.40x
py_globals	124 MB	171 MB: 1.38x	169 MB: 1.36x
event_widgets runaway (*)	188 MB	2,255 MB: 11.97x	2,234 MB: 11.85x
Geometric mean	(ref)	1.39x	1.43x

Benchmarks with tag 'workloads': time

Benchmark	default	FT 64K	FT 16K
web_service	52,525 ops/s	48,112 ops/s: 1.09x slower	51,578 ops/s: 1.02x slower
orm_session mixed	2,181 ops/s	1,759 ops/s: 1.24x slower	1,785 ops/s: 1.22x slower
async_server	8,841 ops/s	7,130 ops/s: 1.24x slower	7,434 ops/s: 1.19x slower
json_churn	64,798 ops/s	57,431 ops/s: 1.13x slower	57,336 ops/s: 1.13x slower
Geometric mean	(ref)	1.17x slower	1.14x slower

Benchmarks with tag 'workloads': peak memory

Benchmark	default	FT 64K	FT 16K
web_service	376 MB	494 MB: 1.31x	502 MB: 1.34x
orm_session mixed	85 MB	124 MB: 1.46x	126 MB: 1.49x
async_server	43 MB	57 MB: 1.34x	55 MB: 1.30x
json_churn	83 MB	98 MB: 1.17x	98 MB: 1.18x
Geometric mean	(ref)	1.32x	1.32x

Benchmarks with tag 'threads': time

Benchmark	default	FT 64K	FT 16K
web_service 4t	61,093 ops/s	148,439 ops/s: 2.43x faster	150,626 ops/s: 2.47x faster
web_service 8t	64,679 ops/s	140,723 ops/s: 2.18x faster	147,049 ops/s: 2.27x faster
orm_session mixed 4t	2,085 ops/s	4,975 ops/s: 2.39x faster	5,266 ops/s: 2.53x faster
Geometric mean	(ref)	2.33x faster	2.42x faster

Benchmarks with tag 'threads': peak memory

Benchmark	default	FT 64K	FT 16K
web_service 4t	399 MB	486 MB: 1.22x	492 MB: 1.23x
web_service 8t	419 MB	492 MB: 1.17x	498 MB: 1.19x
orm_session mixed 4t	107 MB	152 MB: 1.42x	147 MB: 1.37x
Geometric mean	(ref)	1.27x	1.26x

Overall

• Time geometric mean (21 benchmarks, excluding (†)): FT 64K 1.01x faster, FT 16K 1.01x slower
• Time geometric mean incl. (†) (23): FT 64K 1.02x faster, FT 16K 1.06x faster
• (†) bigbuffers: README says the runs are too short to time; shown but excluded from the main geomean. (*) event_widgets runaway, ext_buffers_malloc: large run-to-run variance.
• Memory usage (23 benchmarks): FT 64K 1.36x, FT 16K 1.38x
• Memory usage without (*) benchmarks (21): FT 64K 1.20x, FT 16K 1.21x
• FT 16K vs FT 64K memory: 1.016x (without (*): 1.006x)

[nascheme]

This seems to be a classic time-vs-space tradeoff. Generally, I would lean towards sticking with upstream 64k since it doesn't make a large difference. However, since Donghee did extensive testing and it does seem a small win with no major regressions, this seems safe to change. Another point: if we can port some of the GC optimizations from the default build, doing fewer heap traversals by combining passes, this would likely be a bigger win.

[DinoV]

Nice!