Inline cache for slots

[gvanrossum]

BPO	42927
Nosy	@gvanrossum, @rhettinger, @methane, @serhiy-storchaka, @1st1, @pablogsal, @isidentical
PRs	bpo-42927: Inline cache for slots #24216 bpo-42927: Update the What's new entry for LOAD_ATTR optimizations #24383
Files	result2.zip

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-01-30.02:03:05.447>
created_at = <Date 2021-01-14.04:45:09.122>
labels = ['3.10', 'performance']
title = 'Inline cache for slots'
updated_at = <Date 2021-01-31.22:55:58.106>
user = 'https://github.com/gvanrossum'

bugs.python.org fields:

activity = <Date 2021-01-31.22:55:58.106>
actor = 'pablogsal'
assignee = 'none'
closed = True
closed_date = <Date 2021-01-30.02:03:05.447>
closer = 'gvanrossum'
components = []
creation = <Date 2021-01-14.04:45:09.122>
creator = 'gvanrossum'
dependencies = []
files = ['49744']
hgrepos = []
issue_num = 42927
keywords = ['patch']
message_count = 15.0
messages = ['385063', '385064', '385066', '385076', '385077', '385078', '385079', '385080', '385081', '385091', '385095', '385103', '385124', '385965', '386046']
nosy_count = 7.0
nosy_names = ['gvanrossum', 'rhettinger', 'methane', 'serhiy.storchaka', 'yselivanov', 'pablogsal', 'BTaskaya']
pr_nums = ['24216', '24383']
priority = 'normal'
resolution = 'fixed'
stage = 'commit review'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue42927'
versions = ['Python 3.10']

[gvanrossum]

I've been thinking about Python performance improvements, and I played around with an inline cache enhancement that supports slots. The results on a very simple benchmark look promising (30% speedup) but I'm terrible with our benchmarking tools, and this should be considered *very* carefully before we go ahead with it, since it could potentially pessimize the inline cache for standard attributes.

I'll attach a PR in a minute.

[serhiy-storchaka]

Slot means so many different things in Python. Here it's about data descriptors created when you set __slots__ in the class definition.

It is amazing that so large speed up can be achieved for such base operation.

[pablogsal]

I will try to run today the pyperformance test suite to see if there is any impact on standard attribute access

[pablogsal]

These are the results:

venv ❯ python -m pyperf compare_to json_old/* -G --min-speed=2 --table
+---------------------+--------------------------------------+------------------------------------------+
| Benchmark | 2021-01-14_09-40-master-14cfa325c298 | 2021-01-14_10-39-slotscache-9123a84ad8d4 |
+=====================+======================================+==========================================+
| float | 148 ms | 133 ms: 1.11x faster |
+---------------------+--------------------------------------+------------------------------------------+
| pidigits | 265 ms | 252 ms: 1.05x faster |
+---------------------+--------------------------------------+------------------------------------------+
| scimark_monte_carlo | 130 ms | 126 ms: 1.03x faster |
+---------------------+--------------------------------------+------------------------------------------+
| spectral_norm | 168 ms | 164 ms: 1.03x faster |
+---------------------+--------------------------------------+------------------------------------------+
| logging_silent | 221 ns | 216 ns: 1.02x faster |
+---------------------+--------------------------------------+------------------------------------------+
| sympy_expand | 598 ms | 585 ms: 1.02x faster |
+---------------------+--------------------------------------+------------------------------------------+
| regex_v8 | 31.9 ms | 32.6 ms: 1.02x slower |
+---------------------+--------------------------------------+------------------------------------------+
| deltablue | 8.93 ms | 9.16 ms: 1.03x slower |
+---------------------+--------------------------------------+------------------------------------------+
| unpickle | 17.0 us | 17.7 us: 1.04x slower |
+---------------------+--------------------------------------+------------------------------------------+
| Geometric mean | (ref) | 1.00x faster |
+---------------------+--------------------------------------+------------------------------------------+

Benchmark hidden because not significant (50): sympy_sum, sqlalchemy_imperative, sympy_str, sympy_integrate, dulwich_log, scimark_fft, genshi_text, tornado_http, regex_dna, sqlalchemy_declarative, mako, meteor_contest, unpickle_pure_python, xml_etree_parse, genshi_xml, scimark_sor, sqlite_synth, pickle_pure_python, nbody, pickle_dict, pyflate, regex_effbot, xml_etree_process, logging_simple, python_startup, 2to3, fannkuch, python_startup_no_site, raytrace, go, hexiom, scimark_lu, json_dumps, richards, logging_format, xml_etree_generate, chaos, telco, pickle_list, unpack_sequence, regex_compile, django_template, json_loads, crypto_pyaes, xml_etree_iterparse, unpickle_list, nqueens, chameleon, scimark_sparse_mat_mult, pickle
Ignored benchmarks (1) of json_old/2021-01-14_09-40-master-14cfa325c298.json.gz: pathlib

I uploaded the pyperf result data to bpo as well

[pablogsal]

So it seems that everything is in the noise range except the "float" benchmark that is 1.11x faster

[pablogsal]

Some microbenchmarks:

CURRENT MASTER:

Variable and attribute read access:
4.5 ns read_local
5.8 ns read_nonlocal
7.8 ns read_global
7.8 ns read_builtin
21.2 ns read_classvar_from_class
19.1 ns read_classvar_from_instance
16.4 ns read_instancevar
23.6 ns read_instancevar_slots
20.2 ns read_namedtuple
40.5 ns read_boundmethod

PR 24216

Variable and attribute read access:
4.5 ns read_local
5.8 ns read_nonlocal
7.8 ns read_global
8.2 ns read_builtin
21.4 ns read_classvar_from_class
19.0 ns read_classvar_from_instance
16.6 ns read_instancevar
10.2 ns read_instancevar_slots
20.5 ns read_namedtuple
42.9 ns read_boundmethod

[1st1]

So it seems that everything is in the noise range except the "float" benchmark that is 1.11x faster

Yeah, this is why. https://github.com/python/pyperformance/blob/master/pyperformance/benchmarks/bm_float.py#L12

This is a great result, IMO. I'm +1 to merge this.

[1st1]

Some microbenchmarks:

Can you add a new one read_slots?

[pablogsal]

This is a great result, IMO. I'm +1 to merge this.

Same here. It would be good if Inada-san could confirm the benchmarks.

Can you add a new one read_slots?

That should be "read_instancevar_slots". Or you refer to some other check?

[rhettinger]

+1 Thanks for this.

[gvanrossum]

Thanks for all the positive feedback! What is the next step?

[pablogsal]

What is the next step?

I would say just finishing the PR (making sure that we do not miss some arcane edge case) and updating the what's new for 3.10 :)

[gvanrossum]

I created a benchmark for this, see python/pyperformance#86.

Next I'll add a blurb and then it's off to reviewers.

[gvanrossum]

New changeset 5c5a938 by Guido van Rossum in branch 'master':
bpo-42927: Inline cache for attributes defined with '__slots__' (bpo-24216)
5c5a938

[pablogsal]

New changeset a776da9 by Pablo Galindo in branch 'master':
bpo-42927: Update the What's new entry for LOAD_ATTR optimizations (GH-24383)
a776da9