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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
|
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.1//EN" "http://www.w3.org/TR/xhtml11/DTD/xhtml11.dtd">
<html xmlns="http://www.w3.org/1999/xhtml"><head><title>Design</title><meta name="generator" content="DocBook XSL-NS Stylesheets V1.76.1"/><meta name="keywords" content=" C++ , library , parallel "/><meta name="keywords" content=" ISO C++ , library "/><link rel="home" href="../spine.html" title="The GNU C++ Library"/><link rel="up" href="parallel_mode.html" title="Chapter 18. Parallel Mode"/><link rel="prev" href="bk01pt03ch18s03.html" title="Using"/><link rel="next" href="bk01pt03ch18s05.html" title="Testing"/></head><body><div class="navheader"><table width="100%" summary="Navigation header"><tr><th colspan="3" align="center">Design</th></tr><tr><td align="left"><a accesskey="p" href="bk01pt03ch18s03.html">Prev</a> </td><th width="60%" align="center">Chapter 18. Parallel Mode</th><td align="right"> <a accesskey="n" href="bk01pt03ch18s05.html">Next</a></td></tr></table><hr/></div><div class="section" title="Design"><div class="titlepage"><div><div><h2 class="title"><a id="manual.ext.parallel_mode.design"/>Design</h2></div></div></div><p>
</p><div class="section" title="Interface Basics"><div class="titlepage"><div><div><h3 class="title"><a id="parallel_mode.design.intro"/>Interface Basics</h3></div></div></div><p>
All parallel algorithms are intended to have signatures that are
equivalent to the ISO C++ algorithms replaced. For instance, the
<code class="function">std::adjacent_find</code> function is declared as:
</p><pre class="programlisting">
namespace std
{
template<typename _FIter>
_FIter
adjacent_find(_FIter, _FIter);
}
</pre><p>
Which means that there should be something equivalent for the parallel
version. Indeed, this is the case:
</p><pre class="programlisting">
namespace std
{
namespace __parallel
{
template<typename _FIter>
_FIter
adjacent_find(_FIter, _FIter);
...
}
}
</pre><p>But.... why the ellipses?
</p><p> The ellipses in the example above represent additional overloads
required for the parallel version of the function. These additional
overloads are used to dispatch calls from the ISO C++ function
signature to the appropriate parallel function (or sequential
function, if no parallel functions are deemed worthy), based on either
compile-time or run-time conditions.
</p><p> The available signature options are specific for the different
algorithms/algorithm classes.</p><p> The general view of overloads for the parallel algorithms look like this:
</p><div class="itemizedlist"><ul class="itemizedlist"><li class="listitem"><p>ISO C++ signature</p></li><li class="listitem"><p>ISO C++ signature + sequential_tag argument</p></li><li class="listitem"><p>ISO C++ signature + algorithm-specific tag type
(several signatures)</p></li></ul></div><p> Please note that the implementation may use additional functions
(designated with the <code class="code">_switch</code> suffix) to dispatch from the
ISO C++ signature to the correct parallel version. Also, some of the
algorithms do not have support for run-time conditions, so the last
overload is therefore missing.
</p></div><div class="section" title="Configuration and Tuning"><div class="titlepage"><div><div><h3 class="title"><a id="parallel_mode.design.tuning"/>Configuration and Tuning</h3></div></div></div><div class="section" title="Setting up the OpenMP Environment"><div class="titlepage"><div><div><h4 class="title"><a id="parallel_mode.design.tuning.omp"/>Setting up the OpenMP Environment</h4></div></div></div><p>
Several aspects of the overall runtime environment can be manipulated
by standard OpenMP function calls.
</p><p>
To specify the number of threads to be used for the algorithms globally,
use the function <code class="function">omp_set_num_threads</code>. An example:
</p><pre class="programlisting">
#include <stdlib.h>
#include <omp.h>
int main()
{
// Explicitly set number of threads.
const int threads_wanted = 20;
omp_set_dynamic(false);
omp_set_num_threads(threads_wanted);
// Call parallel mode algorithms.
return 0;
}
</pre><p>
Some algorithms allow the number of threads being set for a particular call,
by augmenting the algorithm variant.
See the next section for further information.
</p><p>
Other parts of the runtime environment able to be manipulated include
nested parallelism (<code class="function">omp_set_nested</code>), schedule kind
(<code class="function">omp_set_schedule</code>), and others. See the OpenMP
documentation for more information.
</p></div><div class="section" title="Compile Time Switches"><div class="titlepage"><div><div><h4 class="title"><a id="parallel_mode.design.tuning.compile"/>Compile Time Switches</h4></div></div></div><p>
To force an algorithm to execute sequentially, even though parallelism
is switched on in general via the macro <code class="constant">_GLIBCXX_PARALLEL</code>,
add <code class="classname">__gnu_parallel::sequential_tag()</code> to the end
of the algorithm's argument list.
</p><p>
Like so:
</p><pre class="programlisting">
std::sort(v.begin(), v.end(), __gnu_parallel::sequential_tag());
</pre><p>
Some parallel algorithm variants can be excluded from compilation by
preprocessor defines. See the doxygen documentation on
<code class="code">compiletime_settings.h</code> and <code class="code">features.h</code> for details.
</p><p>
For some algorithms, the desired variant can be chosen at compile-time by
appending a tag object. The available options are specific to the particular
algorithm (class).
</p><p>
For the "embarrassingly parallel" algorithms, there is only one "tag object
type", the enum _Parallelism.
It takes one of the following values,
<code class="code">__gnu_parallel::parallel_tag</code>,
<code class="code">__gnu_parallel::balanced_tag</code>,
<code class="code">__gnu_parallel::unbalanced_tag</code>,
<code class="code">__gnu_parallel::omp_loop_tag</code>,
<code class="code">__gnu_parallel::omp_loop_static_tag</code>.
This means that the actual parallelization strategy is chosen at run-time.
(Choosing the variants at compile-time will come soon.)
</p><p>
For the following algorithms in general, we have
<code class="code">__gnu_parallel::parallel_tag</code> and
<code class="code">__gnu_parallel::default_parallel_tag</code>, in addition to
<code class="code">__gnu_parallel::sequential_tag</code>.
<code class="code">__gnu_parallel::default_parallel_tag</code> chooses the default
algorithm at compiletime, as does omitting the tag.
<code class="code">__gnu_parallel::parallel_tag</code> postpones the decision to runtime
(see next section).
For all tags, the number of threads desired for this call can optionally be
passed to the respective tag's constructor.
</p><p>
The <code class="code">multiway_merge</code> algorithm comes with the additional choices,
<code class="code">__gnu_parallel::exact_tag</code> and
<code class="code">__gnu_parallel::sampling_tag</code>.
Exact and sampling are the two available splitting strategies.
</p><p>
For the <code class="code">sort</code> and <code class="code">stable_sort</code> algorithms, there are
several additional choices, namely
<code class="code">__gnu_parallel::multiway_mergesort_tag</code>,
<code class="code">__gnu_parallel::multiway_mergesort_exact_tag</code>,
<code class="code">__gnu_parallel::multiway_mergesort_sampling_tag</code>,
<code class="code">__gnu_parallel::quicksort_tag</code>, and
<code class="code">__gnu_parallel::balanced_quicksort_tag</code>.
Multiway mergesort comes with the two splitting strategies for multi-way
merging. The quicksort options cannot be used for <code class="code">stable_sort</code>.
</p></div><div class="section" title="Run Time Settings and Defaults"><div class="titlepage"><div><div><h4 class="title"><a id="parallel_mode.design.tuning.settings"/>Run Time Settings and Defaults</h4></div></div></div><p>
The default parallelization strategy, the choice of specific algorithm
strategy, the minimum threshold limits for individual parallel
algorithms, and aspects of the underlying hardware can be specified as
desired via manipulation
of <code class="classname">__gnu_parallel::_Settings</code> member data.
</p><p>
First off, the choice of parallelization strategy: serial, parallel,
or heuristically deduced. This corresponds
to <code class="code">__gnu_parallel::_Settings::algorithm_strategy</code> and is a
value of enum <span class="type">__gnu_parallel::_AlgorithmStrategy</span>
type. Choices
include: <span class="type">heuristic</span>, <span class="type">force_sequential</span>,
and <span class="type">force_parallel</span>. The default is <span class="type">heuristic</span>.
</p><p>
Next, the sub-choices for algorithm variant, if not fixed at compile-time.
Specific algorithms like <code class="function">find</code> or <code class="function">sort</code>
can be implemented in multiple ways: when this is the case,
a <code class="classname">__gnu_parallel::_Settings</code> member exists to
pick the default strategy. For
example, <code class="code">__gnu_parallel::_Settings::sort_algorithm</code> can
have any values of
enum <span class="type">__gnu_parallel::_SortAlgorithm</span>: <span class="type">MWMS</span>, <span class="type">QS</span>,
or <span class="type">QS_BALANCED</span>.
</p><p>
Likewise for setting the minimal threshold for algorithm
parallelization. Parallelism always incurs some overhead. Thus, it is
not helpful to parallelize operations on very small sets of
data. Because of this, measures are taken to avoid parallelizing below
a certain, pre-determined threshold. For each algorithm, a minimum
problem size is encoded as a variable in the
active <code class="classname">__gnu_parallel::_Settings</code> object. This
threshold variable follows the following naming scheme:
<code class="code">__gnu_parallel::_Settings::[algorithm]_minimal_n</code>. So,
for <code class="function">fill</code>, the threshold variable
is <code class="code">__gnu_parallel::_Settings::fill_minimal_n</code>,
</p><p>
Finally, hardware details like L1/L2 cache size can be hardwired
via <code class="code">__gnu_parallel::_Settings::L1_cache_size</code> and friends.
</p><p>
</p><p>
All these configuration variables can be changed by the user, if
desired.
There exists one global instance of the class <code class="classname">_Settings</code>,
i. e. it is a singleton. It can be read and written by calling
<code class="code">__gnu_parallel::_Settings::get</code> and
<code class="code">__gnu_parallel::_Settings::set</code>, respectively.
Please note that the first call return a const object, so direct manipulation
is forbidden.
See <a class="link" href="http://gcc.gnu.org/onlinedocs/libstdc++/latest-doxygen/a01005.html">
<code class="filename">settings.h</code></a>
for complete details.
</p><p>
A small example of tuning the default:
</p><pre class="programlisting">
#include <parallel/algorithm>
#include <parallel/settings.h>
int main()
{
__gnu_parallel::_Settings s;
s.algorithm_strategy = __gnu_parallel::force_parallel;
__gnu_parallel::_Settings::set(s);
// Do work... all algorithms will be parallelized, always.
return 0;
}
</pre></div></div><div class="section" title="Implementation Namespaces"><div class="titlepage"><div><div><h3 class="title"><a id="parallel_mode.design.impl"/>Implementation Namespaces</h3></div></div></div><p> One namespace contain versions of code that are always
explicitly sequential:
<code class="code">__gnu_serial</code>.
</p><p> Two namespaces contain the parallel mode:
<code class="code">std::__parallel</code> and <code class="code">__gnu_parallel</code>.
</p><p> Parallel implementations of standard components, including
template helpers to select parallelism, are defined in <code class="code">namespace
std::__parallel</code>. For instance, <code class="function">std::transform</code> from <code class="filename">algorithm</code> has a parallel counterpart in
<code class="function">std::__parallel::transform</code> from <code class="filename">parallel/algorithm</code>. In addition, these parallel
implementations are injected into <code class="code">namespace
__gnu_parallel</code> with using declarations.
</p><p> Support and general infrastructure is in <code class="code">namespace
__gnu_parallel</code>.
</p><p> More information, and an organized index of types and functions
related to the parallel mode on a per-namespace basis, can be found in
the generated source documentation.
</p></div></div><div class="navfooter"><hr/><table width="100%" summary="Navigation footer"><tr><td align="left"><a accesskey="p" href="bk01pt03ch18s03.html">Prev</a> </td><td align="center"><a accesskey="u" href="parallel_mode.html">Up</a></td><td align="right"> <a accesskey="n" href="bk01pt03ch18s05.html">Next</a></td></tr><tr><td align="left" valign="top">Using </td><td align="center"><a accesskey="h" href="../spine.html">Home</a></td><td align="right" valign="top"> Testing</td></tr></table></div></body></html>
|
