1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
| # /// script
# dependencies = ["numpy"]
# ///
"""How many bytes does one small Python object cost?
Creates N instances of a few variants of the same 3-attribute class and
measures the memory they add with tracemalloc, plus how fast their
attributes read. Then wide classes, which fall off CPython's fast path, and
no objects at all: the attributes as raw ints in NumPy arrays.
uv run --python 3.13 measure.py
Written for "A million objects: where do the bytes go?":
https://ksh.fyi/blogs/python-slots/
"""
import gc
import sys
import time
import tracemalloc
N = 1_000_000
class Tree:
def __init__(self, x, y, age):
self.x = x
self.y = y
self.age = age
class TreeOwnDict:
"""Opts out of key-sharing: every instance gets its own ordinary dict,
roughly how all instances looked before Python 3.3 (PEP 412)."""
def __init__(self, x, y, age):
self.__dict__ = {"x": x, "y": y, "age": age}
class TreeSlots:
__slots__ = ("x", "y", "age")
def __init__(self, x, y, age):
self.x = x
self.y = y
self.age = age
def make(cls):
objs = [None] * N # allocate the list up front so it isn't counted
gc.collect()
tracemalloc.start()
before = tracemalloc.get_traced_memory()[0]
for i in range(N):
# small ints (< 256) are cached by CPython, so they cost nothing here
objs[i] = cls(i % 100, i % 200, i % 50)
used = tracemalloc.get_traced_memory()[0] - before
tracemalloc.stop()
return objs, used / N
def touch_dict(objs):
"""Reading __dict__ forces 3.11+ to build a real dict object."""
gc.collect()
tracemalloc.start()
before = tracemalloc.get_traced_memory()[0]
for o in objs:
o.__dict__
used = tracemalloc.get_traced_memory()[0] - before
tracemalloc.stop()
return used / N
def read_ns(objs):
start = time.perf_counter()
for o in objs:
o.x
o.y
o.age
return (time.perf_counter() - start) / (3 * N) * 1e9
def age_ns(objs):
"""Age every tree by a year, one object at a time."""
start = time.perf_counter()
for o in objs:
o.age += 1
return (time.perf_counter() - start) / N * 1e9
def numpy_forest():
"""No Tree objects at all: one NumPy array of raw 8-byte ints per attribute."""
import numpy as np
gc.collect()
tracemalloc.start() # NumPy reports its array buffers to tracemalloc
before = tracemalloc.get_traced_memory()[0]
i = np.arange(N)
x, y, age = i % 100, i % 200, i % 50
del i
used = tracemalloc.get_traced_memory()[0] - before
tracemalloc.stop()
start = time.perf_counter()
age += 1
aging = (time.perf_counter() - start) / N * 1e9
start = time.perf_counter()
for k in range(N):
x[k]
y[k]
age[k]
read = (time.perf_counter() - start) / (3 * N) * 1e9
return used / N, aging, read
def wide(n, slots=False):
"""A class that sets n attributes in __init__."""
names = ["a%d" % k for k in range(n)]
def __init__(self, *args):
for name in names:
setattr(self, name, 1)
ns = {"__init__": __init__}
if slots:
ns["__slots__"] = tuple(names)
return type("Wide%d" % n, (), ns)
def main():
version = "%d.%d" % sys.version_info[:2]
rows = []
aging = {}
for cls in (TreeOwnDict, Tree, TreeSlots):
objs, per_obj = make(cls)
rows.append((cls.__name__, per_obj, sys.getsizeof(objs[0]), read_ns(objs)))
aging[cls.__name__] = age_ns(objs)
if cls is Tree:
extra = touch_dict(objs)
rows.append(("Tree, __dict__ read", per_obj + extra,
sys.getsizeof(objs[0]), read_ns(objs)))
del objs
print("Python %s, %s instances" % (version, format(N, ",")))
print("%-22s %10s %10s %10s" % ("variant", "bytes/obj", "getsizeof", "ns/read"))
for name, per_obj, size, ns in rows:
print("%-22s %10.1f %10d %10.1f" % (name, per_obj, size, ns))
print()
print("%-22s %10s" % ("wide classes", "bytes/obj"))
shapes = [("%d attrs" % n, wide(n)) for n in (29, 30)]
shapes += [("30 attrs, slots", wide(30, slots=True))]
for name, cls in shapes:
objs, per_obj = make(cls)
print("%-22s %10.1f" % (name, per_obj))
del objs
print()
per_tree, numpy_aging, numpy_read = numpy_forest()
print("NumPy columns: %.1f bytes/tree, %.1f ns/read" % (per_tree, numpy_read))
print()
print("%-22s %10s" % ("age every tree", "ns/tree"))
for name, ns in aging.items():
print("%-22s %10.2f" % (name + " loop", ns))
print("%-22s %10.2f" % ("NumPy age += 1", numpy_aging))
if __name__ == "__main__":
main()
|