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
|
Performance
-----------
.. currentmodule:: numpy.random
Recommendation
**************
The recommended generator for general use is `PCG64`. It is
statistically high quality, full-featured, and fast on most platforms, but
somewhat slow when compiled for 32-bit processes.
`Philox` is fairly slow, but its statistical properties have
very high quality, and it is easy to get assuredly-independent stream by using
unique keys. If that is the style you wish to use for parallel streams, or you
are porting from another system that uses that style, then
`Philox` is your choice.
`SFC64` is statistically high quality and very fast. However, it
lacks jumpability. If you are not using that capability and want lots of speed,
even on 32-bit processes, this is your choice.
`MT19937` `fails some statistical tests`_ and is not especially
fast compared to modern PRNGs. For these reasons, we mostly do not recommend
using it on its own, only through the legacy `~.RandomState` for
reproducing old results. That said, it has a very long history as a default in
many systems.
.. _`fails some statistical tests`: https://www.iro.umontreal.ca/~lecuyer/myftp/papers/testu01.pdf
Timings
*******
The timings below are the time in ns to produce 1 random value from a
specific distribution. The original `MT19937` generator is
much slower since it requires 2 32-bit values to equal the output of the
faster generators.
Integer performance has a similar ordering.
The pattern is similar for other, more complex generators. The normal
performance of the legacy `RandomState` generator is much
lower than the other since it uses the Box-Muller transformation rather
than the Ziggurat generator. The performance gap for Exponentials is also
large due to the cost of computing the log function to invert the CDF.
The column labeled MT19973 is used the same 32-bit generator as
`RandomState` but produces random values using
`Generator`.
.. csv-table::
:header: ,MT19937,PCG64,Philox,SFC64,RandomState
:widths: 14,14,14,14,14,14
32-bit Unsigned Ints,3.2,2.7,4.9,2.7,3.2
64-bit Unsigned Ints,5.6,3.7,6.3,2.9,5.7
Uniforms,7.3,4.1,8.1,3.1,7.3
Normals,13.1,10.2,13.5,7.8,34.6
Exponentials,7.9,5.4,8.5,4.1,40.3
Gammas,34.8,28.0,34.7,25.1,58.1
Binomials,25.0,21.4,26.1,19.5,25.2
Laplaces,45.1,40.7,45.5,38.1,45.6
Poissons,67.6,52.4,69.2,46.4,78.1
The next table presents the performance in percentage relative to values
generated by the legacy generator, ``RandomState(MT19937())``. The overall
performance was computed using a geometric mean.
.. csv-table::
:header: ,MT19937,PCG64,Philox,SFC64
:widths: 14,14,14,14,14
32-bit Unsigned Ints,101,121,67,121
64-bit Unsigned Ints,102,156,91,199
Uniforms,100,179,90,235
Normals,263,338,257,443
Exponentials,507,752,474,985
Gammas,167,207,167,231
Binomials,101,118,96,129
Laplaces,101,112,100,120
Poissons,116,149,113,168
Overall,144,192,132,225
.. note::
All timings were taken using Linux on a i5-3570 processor.
Performance on different Operating Systems
******************************************
Performance differs across platforms due to compiler and hardware availability
(e.g., register width) differences. The default bit generator has been chosen
to perform well on 64-bit platforms. Performance on 32-bit operating systems
is very different.
The values reported are normalized relative to the speed of MT19937 in
each table. A value of 100 indicates that the performance matches the MT19937.
Higher values indicate improved performance. These values cannot be compared
across tables.
64-bit Linux
~~~~~~~~~~~~
=================== ========= ======= ======== =======
Distribution MT19937 PCG64 Philox SFC64
=================== ========= ======= ======== =======
32-bit Unsigned Int 100 119.8 67.7 120.2
64-bit Unsigned Int 100 152.9 90.8 213.3
Uniforms 100 179.0 87.0 232.0
Normals 100 128.5 99.2 167.8
Exponentials 100 148.3 93.0 189.3
**Overall** 100 144.3 86.8 180.0
=================== ========= ======= ======== =======
64-bit Windows
~~~~~~~~~~~~~~
The relative performance on 64-bit Linux and 64-bit Windows is broadly similar.
=================== ========= ======= ======== =======
Distribution MT19937 PCG64 Philox SFC64
=================== ========= ======= ======== =======
32-bit Unsigned Int 100 129.1 35.0 135.0
64-bit Unsigned Int 100 146.9 35.7 176.5
Uniforms 100 165.0 37.0 192.0
Normals 100 128.5 48.5 158.0
Exponentials 100 151.6 39.0 172.8
**Overall** 100 143.6 38.7 165.7
=================== ========= ======= ======== =======
32-bit Windows
~~~~~~~~~~~~~~
The performance of 64-bit generators on 32-bit Windows is much lower than on 64-bit
operating systems due to register width. MT19937, the generator that has been
in NumPy since 2005, operates on 32-bit integers.
=================== ========= ======= ======== =======
Distribution MT19937 PCG64 Philox SFC64
=================== ========= ======= ======== =======
32-bit Unsigned Int 100 30.5 21.1 77.9
64-bit Unsigned Int 100 26.3 19.2 97.0
Uniforms 100 28.0 23.0 106.0
Normals 100 40.1 31.3 112.6
Exponentials 100 33.7 26.3 109.8
**Overall** 100 31.4 23.8 99.8
=================== ========= ======= ======== =======
.. note::
Linux timings used Ubuntu 18.04 and GCC 7.4. Windows timings were made on
Windows 10 using Microsoft C/C++ Optimizing Compiler Version 19 (Visual
Studio 2015). All timings were produced on a i5-3570 processor.
|
