Reduce lru_cache memory overhead.

[methane]

BPO	32422
Nosy	@rhettinger, @methane, @serhiy-storchaka
PRs	bpo-32422: Reduce lru_cache memory usage. #5008
Files	lru_bench.py

^{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 = 'https://github.com/rhettinger'
closed_at = <Date 2017-12-25.17:03:57.561>
created_at = <Date 2017-12-24.05:19:02.014>
labels = ['interpreter-core', '3.7', 'performance']
title = 'Reduce lru_cache memory overhead.'
updated_at = <Date 2017-12-25.17:03:57.561>
user = 'https://github.com/methane'

bugs.python.org fields:

activity = <Date 2017-12-25.17:03:57.561>
actor = 'methane'
assignee = 'rhettinger'
closed = True
closed_date = <Date 2017-12-25.17:03:57.561>
closer = 'methane'
components = ['Interpreter Core']
creation = <Date 2017-12-24.05:19:02.014>
creator = 'methane'
dependencies = []
files = ['47348']
hgrepos = []
issue_num = 32422
keywords = ['patch']
message_count = 13.0
messages = ['308980', '308981', '308985', '308986', '309003', '309006', '309007', '309013', '309016', '309029', '309031', '309032', '309039']
nosy_count = 3.0
nosy_names = ['rhettinger', 'methane', 'serhiy.storchaka']
pr_nums = ['5008']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'resource usage'
url = 'https://bugs.python.org/issue32422'
versions = ['Python 3.7']

[methane]

Currently, functools.lru_cache implement own doubly-linked list.

But it is inefficient than OrderedDict because each link node is GC object.
So it may eat more memory and take longer GC time.

I added two private C API for OrderedDict and make lru_cache use it.

• _PyODict_LRUGetItem(): Similar to PyDict_GetItemWithHash() + od.move_to_end(last=True).
• _PyODict_PopItem(): Similar to odict.popitem().

Why I didn't implement C version of move_to_end() is to reduce lookup.
_PyODict_LRUGetItem(key) lookup key once while
od[key]; od.move_to_end(key) lookup key twice.

I'll benchmark it and report result here.

[methane]

Current implementation (no news entry yet):
https://github.com/methane/cpython/pull/10/files

[serhiy-storchaka]

This is a duplicate of bpo-28239.

[serhiy-storchaka]

Ah, sorry, you use OrderedDict instead of just ordered dict. It should have different timing and memory consumption.

[methane]

Hmm, it seems my implementation is 30% slower when many mishit scenario.
Maybe, dict is faster than OrderedDict about massive insert/discard. But I need to profile it.

On the other hand, GC speed looks about 2x faster as expected.

$ ./python -m perf compare_to master.json patched.json -G
Slower (5):
- lru_1000_100: 217 ns +- 6 ns -> 302 ns +- 6 ns: 1.39x slower (+39%)
- lru_10000_1000: 225 ns +- 4 ns -> 309 ns +- 2 ns: 1.37x slower (+37%)
- lru_100_1000: 114 ns +- 5 ns -> 119 ns +- 1 ns: 1.05x slower (+5%)
- lru_100_100: 115 ns +- 6 ns -> 119 ns +- 1 ns: 1.03x slower (+3%)
- lru_1000_1000: 134 ns +- 6 ns -> 136 ns +- 1 ns: 1.02x slower (+2%)

Faster (4):

• gc(1000000): 98.3 ms +- 0.3 ms -> 37.9 ms +- 0.2 ms: 2.59x faster (-61%)
• gc(100000): 11.7 ms +- 0.0 ms -> 5.10 ms +- 0.02 ms: 2.29x faster (-56%)
• gc(10000): 1.48 ms +- 0.02 ms -> 1.04 ms +- 0.01 ms: 1.41x faster (-29%)
• lru_10_100: 149 ns +- 6 ns -> 147 ns +- 2 ns: 1.02x faster (-2%)

[methane]

I found odict.pop() and odict.popitem() is very inefficient because
it look up key multiple times.
odict seems not optimized well and very slow than dict in some area...

I'll try to optimize it in holidays.

[rhettinger]

Please stop revising every single thing you look at. The traditional design of LRU caches used doubly linked lists for a reason. In particular, when there is a high hit rate, the links can be updated without churning the underlying dictionary.

[rhettinger]

FWIW, I'm the original author and designer of this code, so it would have been appropriate to assign this to me for sign-off on any proposed changes.

[serhiy-storchaka]

FWIW, I'm the original author and designer of this code, so it would have
been appropriate to assign this to me for sign-off on any proposed changes.

Not me (the C implementation)? ;-)

[methane]

Please stop revising every single thing you look at. The traditional design of LRU caches used doubly linked lists for a reason. In particular, when there is a high hit rate, the links can be updated without churning the underlying dictionary.

I don't proposing removing doubly linked list; OrderedDict uses
doubly-linked list too, and I found no problem for most-hit scenario.

On the other hand, I found problem of OrderedDict for most mis hit
scenario.

Now I think lru_cache's implementation is better OrderedDict.
PyODict is slower than lru_cache's dict + linked list because of
historical reason (compatibility with pure Python implemantation.)

So I stop trying to remove lru_cache's own implementation.
I'll try to reduce overhead of lru_cache, by removing GC header
from link node.