doc/html/mpi/python.html

   1 <html>
   2 <head>
   3 <meta http-equiv="Content-Type" content="text/html; charset=US-ASCII">
   4 <title>Python Bindings</title>
   5 <link rel="stylesheet" href="../../../doc/src/boostbook.css" type="text/css">
   6 <meta name="generator" content="DocBook XSL Stylesheets V1.78.1">
   7 <link rel="home" href="../index.html" title="The Boost C++ Libraries BoostBook Documentation Subset">
   8 <link rel="up" href="../mpi.html" title="Chapter&#160;20.&#160;Boost.MPI">
   9 <link rel="prev" href="../boost/mpi/timer.html" title="Class timer">
  10 <link rel="next" href="../program_options.html" title="Chapter&#160;21.&#160;Boost.Program_options">
  11 </head>
  12 <body bgcolor="white" text="black" link="#0000FF" vlink="#840084" alink="#0000FF">
  13 <table cellpadding="2" width="100%"><tr>
  14 <td valign="top"><img alt="Boost C++ Libraries" width="277" height="86" src="../../../boost.png"></td>
  15 <td align="center"><a href="../../../index.html">Home</a></td>
  16 <td align="center"><a href="../../../libs/libraries.htm">Libraries</a></td>
  17 <td align="center"><a href="http://www.boost.org/users/people.html">People</a></td>
  18 <td align="center"><a href="http://www.boost.org/users/faq.html">FAQ</a></td>
  19 <td align="center"><a href="../../../more/index.htm">More</a></td>
  20 </tr></table>
  21 <hr>
  22 <div class="spirit-nav">
  23 <a accesskey="p" href="../boost/mpi/timer.html"><img src="../../../doc/src/images/prev.png" alt="Prev"></a><a accesskey="u" href="../mpi.html"><img src="../../../doc/src/images/up.png" alt="Up"></a><a accesskey="h" href="../index.html"><img src="../../../doc/src/images/home.png" alt="Home"></a><a accesskey="n" href="../program_options.html"><img src="../../../doc/src/images/next.png" alt="Next"></a>
  24 </div>
  25 <div class="section">
  26 <div class="titlepage"><div><div><h2 class="title" style="clear: both">
  27 <a name="mpi.python"></a>Python Bindings</h2></div></div></div>
  28 <div class="toc"><dl class="toc">
  29 <dt><span class="section"><a href="python.html#mpi.python_quickstart">Quickstart</a></span></dt>
  30 <dt><span class="section"><a href="python.html#mpi.python_user_data">Transmitting User-Defined Data</a></span></dt>
  31 <dt><span class="section"><a href="python.html#mpi.python_collectives">Collectives</a></span></dt>
  32 <dt><span class="section"><a href="python.html#mpi.python_skeleton_content">Skeleton/Content Mechanism</a></span></dt>
  33 <dt><span class="section"><a href="python.html#mpi.design">Design Philosophy</a></span></dt>
  34 <dt><span class="section"><a href="python.html#mpi.threading">Threads</a></span></dt>
  35 <dt><span class="section"><a href="python.html#mpi.performance">Performance Evaluation</a></span></dt>
  36 <dt><span class="section"><a href="python.html#mpi.history">Revision History</a></span></dt>
  37 <dt><span class="section"><a href="python.html#mpi.acknowledge">Acknowledgments</a></span></dt>
  38 </dl></div>
  39 <p>
  40       Boost.MPI provides an alternative MPI interface from the <a href="http://www.python.org" target="_top">Python</a>
  41       programming language via the <code class="computeroutput"><span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span></code> module.
  42       The Boost.MPI Python bindings, built on top of the C++ Boost.MPI using the
  43       <a href="http://www.boost.org/libs/python/doc" target="_top">Boost.Python</a> library,
  44       provide nearly all of the functionality of Boost.MPI within a dynamic, object-oriented
  45       language.
  46     </p>
  47 <p>
  48       The Boost.MPI Python module can be built and installed from the <code class="computeroutput"><span class="identifier">libs</span><span class="special">/</span><span class="identifier">mpi</span><span class="special">/</span><span class="identifier">build</span></code> directory.
  49       Just follow the <a class="link" href="getting_started.html#mpi.config" title="Configure and Build">configuration</a> and <a class="link" href="getting_started.html#mpi.installation" title="Installing and Using Boost.MPI">installation</a>
  50       instructions for the C++ Boost.MPI. Once you have installed the Python module,
  51       be sure that the installation location is in your <code class="computeroutput"><span class="identifier">PYTHONPATH</span></code>.
  52     </p>
  53 <div class="section">
  54 <div class="titlepage"><div><div><h3 class="title">
  55 <a name="mpi.python_quickstart"></a>Quickstart</h3></div></div></div>
  56 <p>
  57         Getting started with the Boost.MPI Python module is as easy as importing
  58         <code class="computeroutput"><span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span></code>. Our first "Hello, World!"
  59         program is just two lines long:
  60       </p>
  61 <pre class="programlisting"><span class="keyword">import</span> <span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span> <span class="keyword">as</span> <span class="identifier">mpi</span>
  62 <span class="keyword">print</span> <span class="string">"I am process %d of %d."</span> <span class="special">%</span> <span class="special">(</span><span class="identifier">mpi</span><span class="special">.</span><span class="identifier">rank</span><span class="special">,</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">size</span><span class="special">)</span>
  63 </pre>
  64 <p>
  65         Go ahead and run this program with several processes. Be sure to invoke the
  66         <code class="computeroutput"><span class="identifier">python</span></code> interpreter from
  67         <code class="computeroutput"><span class="identifier">mpirun</span></code>, e.g.,
  68       </p>
  69 <pre class="programlisting">mpirun -np 5 python hello_world.py
  70 </pre>
  71 <p>
  72         This will return output such as:
  73       </p>
  74 <pre class="programlisting">I am process 1 of 5.
  75 I am process 3 of 5.
  76 I am process 2 of 5.
  77 I am process 4 of 5.
  78 I am process 0 of 5.
  79 </pre>
  80 <p>
  81         Point-to-point operations in Boost.MPI have nearly the same syntax in Python
  82         as in C++. We can write a simple two-process Python program that prints "Hello,
  83         world!" by transmitting Python strings:
  84       </p>
  85 <pre class="programlisting"><span class="keyword">import</span> <span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span> <span class="keyword">as</span> <span class="identifier">mpi</span>
  86
  87 <span class="keyword">if</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">rank</span> <span class="special">==</span> <span class="number">0</span><span class="special">:</span>
  88   <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">send</span><span class="special">(</span><span class="number">1</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="string">'Hello'</span><span class="special">)</span>
  89   <span class="identifier">msg</span> <span class="special">=</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">recv</span><span class="special">(</span><span class="number">1</span><span class="special">,</span> <span class="number">1</span><span class="special">)</span>
  90   <span class="keyword">print</span> <span class="identifier">msg</span><span class="special">,</span><span class="string">'!'</span>
  91 <span class="keyword">else</span><span class="special">:</span>
  92   <span class="identifier">msg</span> <span class="special">=</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">recv</span><span class="special">(</span><span class="number">0</span><span class="special">,</span> <span class="number">0</span><span class="special">)</span>
  93   <span class="keyword">print</span> <span class="special">(</span><span class="identifier">msg</span> <span class="special">+</span> <span class="string">', '</span><span class="special">),</span>
  94   <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">send</span><span class="special">(</span><span class="number">0</span><span class="special">,</span> <span class="number">1</span><span class="special">,</span> <span class="string">'world'</span><span class="special">)</span>
  95 </pre>
  96 <p>
  97         There are only a few notable differences between this Python code and the
  98         example <a class="link" href="tutorial.html#mpi.point_to_point" title="Point-to-Point communication">in the C++ tutorial</a>. First
  99         of all, we don't need to write any initialization code in Python: just loading
 100         the <code class="computeroutput"><span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span></code> module makes the appropriate <code class="computeroutput"><span class="identifier">MPI_Init</span></code> and <code class="computeroutput"><span class="identifier">MPI_Finalize</span></code>
 101         calls. Second, we're passing Python objects from one process to another through
 102         MPI. Any Python object that can be pickled can be transmitted; the next section
 103         will describe in more detail how the Boost.MPI Python layer transmits objects.
 104         Finally, when we receive objects with <code class="computeroutput"><span class="identifier">recv</span></code>,
 105         we don't need to specify the type because transmission of Python objects
 106         is polymorphic.
 107       </p>
 108 <p>
 109         When experimenting with Boost.MPI in Python, don't forget that help is always
 110         available via <code class="computeroutput"><span class="identifier">pydoc</span></code>: just
 111         pass the name of the module or module entity on the command line (e.g.,
 112         <code class="computeroutput"><span class="identifier">pydoc</span> <span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span><span class="special">.</span><span class="identifier">communicator</span></code>) to receive complete reference
 113         documentation. When in doubt, try it!
 114       </p>
 115 </div>
 116 <div class="section">
 117 <div class="titlepage"><div><div><h3 class="title">
 118 <a name="mpi.python_user_data"></a>Transmitting User-Defined Data</h3></div></div></div>
 119 <p>
 120         Boost.MPI can transmit user-defined data in several different ways. Most
 121         importantly, it can transmit arbitrary <a href="http://www.python.org" target="_top">Python</a>
 122         objects by pickling them at the sender and unpickling them at the receiver,
 123         allowing arbitrarily complex Python data structures to interoperate with
 124         MPI.
 125       </p>
 126 <p>
 127         Boost.MPI also supports efficient serialization and transmission of C++ objects
 128         (that have been exposed to Python) through its C++ interface. Any C++ type
 129         that provides (de-)serialization routines that meet the requirements of the
 130         Boost.Serialization library is eligible for this optimization, but the type
 131         must be registered in advance. To register a C++ type, invoke the C++ function
 132         <code class="computeroutput"><a class="link" href="../boost/mpi/python/register_serialized.html" title="Function template register_serialized">register_serialized</a></code>.
 133         If your C++ types come from other Python modules (they probably will!), those
 134         modules will need to link against the <code class="computeroutput"><span class="identifier">boost_mpi</span></code>
 135         and <code class="computeroutput"><span class="identifier">boost_mpi_python</span></code> libraries
 136         as described in the <a class="link" href="getting_started.html#mpi.installation" title="Installing and Using Boost.MPI">installation section</a>.
 137         Note that you do <span class="bold"><strong>not</strong></span> need to link against
 138         the Boost.MPI Python extension module.
 139       </p>
 140 <p>
 141         Finally, Boost.MPI supports separation of the structure of an object from
 142         the data it stores, allowing the two pieces to be transmitted separately.
 143         This "skeleton/content" mechanism, described in more detail in
 144         a later section, is a communication optimization suitable for problems with
 145         fixed data structures whose internal data changes frequently.
 146       </p>
 147 </div>
 148 <div class="section">
 149 <div class="titlepage"><div><div><h3 class="title">
 150 <a name="mpi.python_collectives"></a>Collectives</h3></div></div></div>
 151 <p>
 152         Boost.MPI supports all of the MPI collectives (<code class="computeroutput"><span class="identifier">scatter</span></code>,
 153         <code class="computeroutput"><span class="identifier">reduce</span></code>, <code class="computeroutput"><span class="identifier">scan</span></code>,
 154         <code class="computeroutput"><span class="identifier">broadcast</span></code>, etc.) for any
 155         type of data that can be transmitted with the point-to-point communication
 156         operations. For the MPI collectives that require a user-specified operation
 157         (e.g., <code class="computeroutput"><span class="identifier">reduce</span></code> and <code class="computeroutput"><span class="identifier">scan</span></code>), the operation can be an arbitrary
 158         Python function. For instance, one could concatenate strings with <code class="computeroutput"><span class="identifier">all_reduce</span></code>:
 159       </p>
 160 <pre class="programlisting"><span class="identifier">mpi</span><span class="special">.</span><span class="identifier">all_reduce</span><span class="special">(</span><span class="identifier">my_string</span><span class="special">,</span> <span class="keyword">lambda</span> <span class="identifier">x</span><span class="special">,</span><span class="identifier">y</span><span class="special">:</span> <span class="identifier">x</span> <span class="special">+</span> <span class="identifier">y</span><span class="special">)</span>
 161 </pre>
 162 <p>
 163         The following module-level functions implement MPI collectives: all_gather
 164         Gather the values from all processes. all_reduce Combine the results from
 165         all processes. all_to_all Every process sends data to every other process.
 166         broadcast Broadcast data from one process to all other processes. gather
 167         Gather the values from all processes to the root. reduce Combine the results
 168         from all processes to the root. scan Prefix reduction of the values from
 169         all processes. scatter Scatter the values stored at the root to all processes.
 170       </p>
 171 </div>
 172 <div class="section">
 173 <div class="titlepage"><div><div><h3 class="title">
 174 <a name="mpi.python_skeleton_content"></a>Skeleton/Content Mechanism</h3></div></div></div>
 175 <div class="toc"><dl class="toc">
 176 <dt><span class="section"><a href="python.html#mpi.python_compatbility">C++/Python MPI Compatibility</a></span></dt>
 177 <dt><span class="section"><a href="python.html#mpi.pythonref">Reference</a></span></dt>
 178 </dl></div>
 179 <p>
 180         Boost.MPI provides a skeleton/content mechanism that allows the transfer
 181         of large data structures to be split into two separate stages, with the skeleton
 182         (or, "shape") of the data structure sent first and the content
 183         (or, "data") of the data structure sent later, potentially several
 184         times, so long as the structure has not changed since the skeleton was transferred.
 185         The skeleton/content mechanism can improve performance when the data structure
 186         is large and its shape is fixed, because while the skeleton requires serialization
 187         (it has an unknown size), the content transfer is fixed-size and can be done
 188         without extra copies.
 189       </p>
 190 <p>
 191         To use the skeleton/content mechanism from Python, you must first register
 192         the type of your data structure with the skeleton/content mechanism <span class="bold"><strong>from C++</strong></span>. The registration function is <code class="computeroutput"><a class="link" href="../boost/mpi/python/register_skel_idp159037008.html" title="Function template register_skeleton_and_content">register_skeleton_and_content</a></code>
 193         and resides in the <code class="computeroutput"><a class="link" href="reference.html#header.boost.mpi.python_hpp" title="Header &lt;boost/mpi/python.hpp&gt;">&lt;boost/mpi/python.hpp&gt;</a></code>
 194         header.
 195       </p>
 196 <p>
 197         Once you have registered your C++ data structures, you can extract the skeleton
 198         for an instance of that data structure with <code class="computeroutput"><span class="identifier">skeleton</span><span class="special">()</span></code>. The resulting <code class="computeroutput"><span class="identifier">skeleton_proxy</span></code>
 199         can be transmitted via the normal send routine, e.g.,
 200       </p>
 201 <pre class="programlisting"><span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">send</span><span class="special">(</span><span class="number">1</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="identifier">skeleton</span><span class="special">(</span><span class="identifier">my_data_structure</span><span class="special">))</span>
 202 </pre>
 203 <p>
 204         <code class="computeroutput"><span class="identifier">skeleton_proxy</span></code> objects can
 205         be received on the other end via <code class="computeroutput"><span class="identifier">recv</span><span class="special">()</span></code>, which stores a newly-created instance
 206         of your data structure with the same "shape" as the sender in its
 207         <code class="computeroutput"><span class="string">"object"</span></code> attribute:
 208       </p>
 209 <pre class="programlisting"><span class="identifier">shape</span> <span class="special">=</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">recv</span><span class="special">(</span><span class="number">0</span><span class="special">,</span> <span class="number">0</span><span class="special">)</span>
 210 <span class="identifier">my_data_structure</span> <span class="special">=</span> <span class="identifier">shape</span><span class="special">.</span><span class="identifier">object</span>
 211 </pre>
 212 <p>
 213         Once the skeleton has been transmitted, the content (accessed via <code class="computeroutput"><span class="identifier">get_content</span></code>) can be transmitted in much
 214         the same way. Note, however, that the receiver also specifies <code class="computeroutput"><span class="identifier">get_content</span><span class="special">(</span><span class="identifier">my_data_structure</span><span class="special">)</span></code>
 215         in its call to receive:
 216       </p>
 217 <pre class="programlisting"><span class="keyword">if</span> <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">rank</span> <span class="special">==</span> <span class="number">0</span><span class="special">:</span>
 218   <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">send</span><span class="special">(</span><span class="number">1</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="identifier">get_content</span><span class="special">(</span><span class="identifier">my_data_structure</span><span class="special">))</span>
 219 <span class="keyword">else</span><span class="special">:</span>
 220   <span class="identifier">mpi</span><span class="special">.</span><span class="identifier">world</span><span class="special">.</span><span class="identifier">recv</span><span class="special">(</span><span class="number">0</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="identifier">get_content</span><span class="special">(</span><span class="identifier">my_data_structure</span><span class="special">))</span>
 221 </pre>
 222 <p>
 223         Of course, this transmission of content can occur repeatedly, if the values
 224         in the data structure--but not its shape--changes.
 225       </p>
 226 <p>
 227         The skeleton/content mechanism is a structured way to exploit the interaction
 228         between custom-built MPI datatypes and <code class="computeroutput"><span class="identifier">MPI_BOTTOM</span></code>,
 229         to eliminate extra buffer copies.
 230       </p>
 231 <div class="section">
 232 <div class="titlepage"><div><div><h4 class="title">
 233 <a name="mpi.python_compatbility"></a>C++/Python MPI Compatibility</h4></div></div></div>
 234 <p>
 235           Boost.MPI is a C++ library whose facilities have been exposed to Python
 236           via the Boost.Python library. Since the Boost.MPI Python bindings are build
 237           directly on top of the C++ library, and nearly every feature of C++ library
 238           is available in Python, hybrid C++/Python programs using Boost.MPI can
 239           interact, e.g., sending a value from Python but receiving that value in
 240           C++ (or vice versa). However, doing so requires some care. Because Python
 241           objects are dynamically typed, Boost.MPI transfers type information along
 242           with the serialized form of the object, so that the object can be received
 243           even when its type is not known. This mechanism differs from its C++ counterpart,
 244           where the static types of transmitted values are always known.
 245         </p>
 246 <p>
 247           The only way to communicate between the C++ and Python views on Boost.MPI
 248           is to traffic entirely in Python objects. For Python, this is the normal
 249           state of affairs, so nothing will change. For C++, this means sending and
 250           receiving values of type <code class="computeroutput"><span class="identifier">boost</span><span class="special">::</span><span class="identifier">python</span><span class="special">::</span><span class="identifier">object</span></code>,
 251           from the <a href="http://www.boost.org/libs/python/doc" target="_top">Boost.Python</a>
 252           library. For instance, say we want to transmit an integer value from Python:
 253         </p>
 254 <pre class="programlisting"><span class="identifier">comm</span><span class="special">.</span><span class="identifier">send</span><span class="special">(</span><span class="number">1</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="number">17</span><span class="special">)</span>
 255 </pre>
 256 <p>
 257           In C++, we would receive that value into a Python object and then <code class="computeroutput"><span class="identifier">extract</span></code> an integer value:
 258         </p>
 259 <pre class="programlisting"><span class="identifier">boost</span><span class="special">::</span><span class="identifier">python</span><span class="special">::</span><span class="identifier">object</span> <span class="identifier">value</span><span class="special">;</span>
 260 <span class="identifier">comm</span><span class="special">.</span><span class="identifier">recv</span><span class="special">(</span><span class="number">0</span><span class="special">,</span> <span class="number">0</span><span class="special">,</span> <span class="identifier">value</span><span class="special">);</span>
 261 <span class="keyword">int</span> <span class="identifier">int_value</span> <span class="special">=</span> <span class="identifier">boost</span><span class="special">::</span><span class="identifier">python</span><span class="special">::</span><span class="identifier">extract</span><span class="special">&lt;</span><span class="keyword">int</span><span class="special">&gt;(</span><span class="identifier">value</span><span class="special">);</span>
 262 </pre>
 263 <p>
 264           In the future, Boost.MPI will be extended to allow improved interoperability
 265           with the C++ Boost.MPI and the C MPI bindings.
 266         </p>
 267 </div>
 268 <div class="section">
 269 <div class="titlepage"><div><div><h4 class="title">
 270 <a name="mpi.pythonref"></a>Reference</h4></div></div></div>
 271 <p>
 272           The Boost.MPI Python module, <code class="computeroutput"><span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span></code>,
 273           has its own <a href="../boost.mpi.html" target="_top">reference documentation</a>,
 274           which is also available using <code class="computeroutput"><span class="identifier">pydoc</span></code>
 275           (from the command line) or <code class="computeroutput"><span class="identifier">help</span><span class="special">(</span><span class="identifier">boost</span><span class="special">.</span><span class="identifier">mpi</span><span class="special">)</span></code> (from the Python interpreter).
 276         </p>
 277 </div>
 278 </div>
 279 <div class="section">
 280 <div class="titlepage"><div><div><h3 class="title">
 281 <a name="mpi.design"></a>Design Philosophy</h3></div></div></div>
 282 <p>
 283         The design philosophy of the Parallel MPI library is very simple: be both
 284         convenient and efficient. MPI is a library built for high-performance applications,
 285         but it's FORTRAN-centric, performance-minded design makes it rather inflexible
 286         from the C++ point of view: passing a string from one process to another
 287         is inconvenient, requiring several messages and explicit buffering; passing
 288         a container of strings from one process to another requires an extra level
 289         of manual bookkeeping; and passing a map from strings to containers of strings
 290         is positively infuriating. The Parallel MPI library allows all of these data
 291         types to be passed using the same simple <code class="computeroutput"><span class="identifier">send</span><span class="special">()</span></code> and <code class="computeroutput"><span class="identifier">recv</span><span class="special">()</span></code> primitives. Likewise, collective operations
 292         such as <code class="computeroutput"><a class="link" href="../boost/mpi/reduce.html" title="Function reduce">reduce()</a></code> allow arbitrary data types
 293         and function objects, much like the C++ Standard Library would.
 294       </p>
 295 <p>
 296         The higher-level abstractions provided for convenience must not have an impact
 297         on the performance of the application. For instance, sending an integer via
 298         <code class="computeroutput"><span class="identifier">send</span></code> must be as efficient
 299         as a call to <code class="computeroutput"><span class="identifier">MPI_Send</span></code>, which
 300         means that it must be implemented by a simple call to <code class="computeroutput"><span class="identifier">MPI_Send</span></code>;
 301         likewise, an integer <code class="computeroutput"><a class="link" href="../boost/mpi/reduce.html" title="Function reduce">reduce()</a></code>
 302         using <code class="computeroutput"><span class="identifier">std</span><span class="special">::</span><span class="identifier">plus</span><span class="special">&lt;</span><span class="keyword">int</span><span class="special">&gt;</span></code> must
 303         be implemented with a call to <code class="computeroutput"><span class="identifier">MPI_Reduce</span></code>
 304         on integers using the <code class="computeroutput"><span class="identifier">MPI_SUM</span></code>
 305         operation: anything less will impact performance. In essence, this is the
 306         "don't pay for what you don't use" principle: if the user is not
 307         transmitting strings, s/he should not pay the overhead associated with strings.
 308       </p>
 309 <p>
 310         Sometimes, achieving maximal performance means foregoing convenient abstractions
 311         and implementing certain functionality using lower-level primitives. For
 312         this reason, it is always possible to extract enough information from the
 313         abstractions in Boost.MPI to minimize the amount of effort required to interface
 314         between Boost.MPI and the C MPI library.
 315       </p>
 316 </div>
 317 <div class="section">
 318 <div class="titlepage"><div><div><h3 class="title">
 319 <a name="mpi.threading"></a>Threads</h3></div></div></div>
 320 <p>
 321         There are an increasing number of hybrid parrallel applications that mix
 322         distributed and shared memory parallelism. To know how to support that model,
 323         one need to know what level of threading support is guaranteed by the MPI
 324         implementation. There are 4 ordered level of possible threading support described
 325         by <code class="computeroutput">mpi::threading::level</code>.
 326         At the lowest level, you should not use threads at all, at the highest level,
 327         any thread can perform MPI call.
 328       </p>
 329 <p>
 330         If you want to use multi-threading in your MPI application, you should indicate
 331         in the environment constructor your preffered threading support. Then probe
 332         the one the librarie did provide, and decide what you can do with it (it
 333         could be nothing, then aborting is a valid option):
 334       </p>
 335 <pre class="programlisting"><span class="preprocessor">#include</span> <span class="special">&lt;</span><span class="identifier">boost</span><span class="special">/</span><span class="identifier">mpi</span><span class="special">/</span><span class="identifier">environment</span><span class="special">.</span><span class="identifier">hpp</span><span class="special">&gt;</span>
 336 <span class="preprocessor">#include</span> <span class="special">&lt;</span><span class="identifier">boost</span><span class="special">/</span><span class="identifier">mpi</span><span class="special">/</span><span class="identifier">communicator</span><span class="special">.</span><span class="identifier">hpp</span><span class="special">&gt;</span>
 337 <span class="preprocessor">#include</span> <span class="special">&lt;</span><span class="identifier">iostream</span><span class="special">&gt;</span>
 338 <span class="keyword">namespace</span> <span class="identifier">mpi</span> <span class="special">=</span> <span class="identifier">boost</span><span class="special">::</span><span class="identifier">mpi</span><span class="special">;</span>
 339 <span class="keyword">namespace</span> <span class="identifier">mt</span>  <span class="special">=</span> <span class="identifier">mpi</span><span class="special">::</span><span class="identifier">threading</span><span class="special">;</span>
 340
 341 <span class="keyword">int</span> <span class="identifier">main</span><span class="special">()</span>
 342 <span class="special">{</span>
 343   <span class="identifier">mpi</span><span class="special">::</span><span class="identifier">environment</span> <span class="identifier">env</span><span class="special">(</span><span class="identifier">mt</span><span class="special">::</span><span class="identifier">funneled</span><span class="special">);</span>
 344   <span class="keyword">if</span> <span class="special">(</span><span class="identifier">env</span><span class="special">.</span><span class="identifier">thread_level</span><span class="special">()</span> <span class="special">&lt;</span> <span class="identifier">mt</span><span class="special">::</span><span class="identifier">funneled</span><span class="special">)</span> <span class="special">{</span>
 345      <span class="identifier">env</span><span class="special">.</span><span class="identifier">abort</span><span class="special">(-</span><span class="number">1</span><span class="special">);</span>
 346   <span class="special">}</span>
 347   <span class="identifier">mpi</span><span class="special">::</span><span class="identifier">communicator</span> <span class="identifier">world</span><span class="special">;</span>
 348   <span class="identifier">std</span><span class="special">::</span><span class="identifier">cout</span> <span class="special">&lt;&lt;</span> <span class="string">"I am process "</span> <span class="special">&lt;&lt;</span> <span class="identifier">world</span><span class="special">.</span><span class="identifier">rank</span><span class="special">()</span> <span class="special">&lt;&lt;</span> <span class="string">" of "</span> <span class="special">&lt;&lt;</span> <span class="identifier">world</span><span class="special">.</span><span class="identifier">size</span><span class="special">()</span>
 349             <span class="special">&lt;&lt;</span> <span class="string">"."</span> <span class="special">&lt;&lt;</span> <span class="identifier">std</span><span class="special">::</span><span class="identifier">endl</span><span class="special">;</span>
 350   <span class="keyword">return</span> <span class="number">0</span><span class="special">;</span>
 351 <span class="special">}</span>
 352 </pre>
 353 </div>
 354 <div class="section">
 355 <div class="titlepage"><div><div><h3 class="title">
 356 <a name="mpi.performance"></a>Performance Evaluation</h3></div></div></div>
 357 <p>
 358         Message-passing performance is crucial in high-performance distributed computing.
 359         To evaluate the performance of Boost.MPI, we modified the standard <a href="http://www.scl.ameslab.gov/netpipe/" target="_top">NetPIPE</a> benchmark (version
 360         3.6.2) to use Boost.MPI and compared its performance against raw MPI. We
 361         ran five different variants of the NetPIPE benchmark:
 362       </p>
 363 <div class="orderedlist"><ol class="orderedlist" type="1">
 364 <li class="listitem">
 365             MPI: The unmodified NetPIPE benchmark.
 366           </li>
 367 <li class="listitem">
 368             Boost.MPI: NetPIPE modified to use Boost.MPI calls for communication.
 369           </li>
 370 <li class="listitem">
 371             MPI (Datatypes): NetPIPE modified to use a derived datatype (which itself
 372             contains a single <code class="computeroutput"><span class="identifier">MPI_BYTE</span></code>)
 373             rathan than a fundamental datatype.
 374           </li>
 375 <li class="listitem">
 376             Boost.MPI (Datatypes): NetPIPE modified to use a user-defined type <code class="computeroutput"><span class="identifier">Char</span></code> in place of the fundamental <code class="computeroutput"><span class="keyword">char</span></code> type. The <code class="computeroutput"><span class="identifier">Char</span></code>
 377             type contains a single <code class="computeroutput"><span class="keyword">char</span></code>,
 378             a <code class="computeroutput"><span class="identifier">serialize</span><span class="special">()</span></code>
 379             method to make it serializable, and specializes <code class="computeroutput"><a class="link" href="../boost/mpi/is_mpi_datatype.html" title="Struct template is_mpi_datatype">is_mpi_datatype</a></code>
 380             to force Boost.MPI to build a derived MPI data type for it.
 381           </li>
 382 <li class="listitem">
 383             Boost.MPI (Serialized): NetPIPE modified to use a user-defined type
 384             <code class="computeroutput"><span class="identifier">Char</span></code> in place of the
 385             fundamental <code class="computeroutput"><span class="keyword">char</span></code> type. This
 386             <code class="computeroutput"><span class="identifier">Char</span></code> type contains a
 387             single <code class="computeroutput"><span class="keyword">char</span></code> and is serializable.
 388             Unlike the Datatypes case, <code class="computeroutput"><a class="link" href="../boost/mpi/is_mpi_datatype.html" title="Struct template is_mpi_datatype">is_mpi_datatype</a></code>
 389             is <span class="bold"><strong>not</strong></span> specialized, forcing Boost.MPI
 390             to perform many, many serialization calls.
 391           </li>
 392 </ol></div>
 393 <p>
 394         The actual tests were performed on the Odin cluster in the <a href="http://www.cs.indiana.edu/" target="_top">Department
 395         of Computer Science</a> at <a href="http://www.iub.edu" target="_top">Indiana University</a>,
 396         which contains 128 nodes connected via Infiniband. Each node contains 4GB
 397         memory and two AMD Opteron processors. The NetPIPE benchmarks were compiled
 398         with Intel's C++ Compiler, version 9.0, Boost 1.35.0 (prerelease), and <a href="http://www.open-mpi.org/" target="_top">Open MPI</a> version 1.1. The NetPIPE
 399         results follow:
 400       </p>
 401 <p>
 402         <span class="inlinemediaobject"><img src="../../../libs/mpi/doc/netpipe.png" alt="netpipe"></span>
 403       </p>
 404 <p>
 405         There are a some observations we can make about these NetPIPE results. First
 406         of all, the top two plots show that Boost.MPI performs on par with MPI for
 407         fundamental types. The next two plots show that Boost.MPI performs on par
 408         with MPI for derived data types, even though Boost.MPI provides a much more
 409         abstract, completely transparent approach to building derived data types
 410         than raw MPI. Overall performance for derived data types is significantly
 411         worse than for fundamental data types, but the bottleneck is in the underlying
 412         MPI implementation itself. Finally, when forcing Boost.MPI to serialize characters
 413         individually, performance suffers greatly. This particular instance is the
 414         worst possible case for Boost.MPI, because we are serializing millions of
 415         individual characters. Overall, the additional abstraction provided by Boost.MPI
 416         does not impair its performance.
 417       </p>
 418 </div>
 419 <div class="section">
 420 <div class="titlepage"><div><div><h3 class="title">
 421 <a name="mpi.history"></a>Revision History</h3></div></div></div>
 422 <div class="itemizedlist"><ul class="itemizedlist" style="list-style-type: disc; ">
 423 <li class="listitem">
 424             <span class="bold"><strong>Boost 1.36.0</strong></span>:
 425             <div class="itemizedlist"><ul class="itemizedlist" style="list-style-type: circle; "><li class="listitem">
 426                   Support for non-blocking operations in Python, from Andreas Kl&#246;ckner
 427                 </li></ul></div>
 428           </li>
 429 <li class="listitem">
 430             <span class="bold"><strong>Boost 1.35.0</strong></span>: Initial release, containing
 431             the following post-review changes
 432             <div class="itemizedlist"><ul class="itemizedlist" style="list-style-type: circle; ">
 433 <li class="listitem">
 434                   Support for arrays in all collective operations
 435                 </li>
 436 <li class="listitem">
 437                   Support default-construction of <code class="computeroutput"><a class="link" href="../boost/mpi/environment.html" title="Class environment">environment</a></code>
 438                 </li>
 439 </ul></div>
 440           </li>
 441 <li class="listitem">
 442             <span class="bold"><strong>2006-09-21</strong></span>: Boost.MPI accepted into
 443             Boost.
 444           </li>
 445 </ul></div>
 446 </div>
 447 <div class="section">
 448 <div class="titlepage"><div><div><h3 class="title">
 449 <a name="mpi.acknowledge"></a>Acknowledgments</h3></div></div></div>
 450 <p>
 451         Boost.MPI was developed with support from Zurcher Kantonalbank. Daniel Egloff
 452         and Michael Gauckler contributed many ideas to Boost.MPI's design, particularly
 453         in the design of its abstractions for MPI data types and the novel skeleton/context
 454         mechanism for large data structures. Prabhanjan (Anju) Kambadur developed
 455         the predecessor to Boost.MPI that proved the usefulness of the Serialization
 456         library in an MPI setting and the performance benefits of specialization
 457         in a C++ abstraction layer for MPI. Jeremy Siek managed the formal review
 458         of Boost.MPI.
 459       </p>
 460 </div>
 461 </div>
 462 <table xmlns:rev="http://www.cs.rpi.edu/~gregod/boost/tools/doc/revision" width="100%"><tr>
 463 <td align="left"></td>
 464 <td align="right"><div class="copyright-footer">Copyright &#169; 2005-2007 Douglas Gregor,
 465       Matthias Troyer, Trustees of Indiana University<p>
 466         Distributed under the Boost Software License, Version 1.0. (See accompanying
 467         file LICENSE_1_0.txt or copy at <a href="http://www.boost.org/LICENSE_1_0.txt" target="_top">
 468         http://www.boost.org/LICENSE_1_0.txt </a>)
 469       </p>
 470 </div></td>
 471 </tr></table>
 472 <hr>
 473 <div class="spirit-nav">
 474 <a accesskey="p" href="../boost/mpi/timer.html"><img src="../../../doc/src/images/prev.png" alt="Prev"></a><a accesskey="u" href="../mpi.html"><img src="../../../doc/src/images/up.png" alt="Up"></a><a accesskey="h" href="../index.html"><img src="../../../doc/src/images/home.png" alt="Home"></a><a accesskey="n" href="../program_options.html"><img src="../../../doc/src/images/next.png" alt="Next"></a>
 475 </div>
 476 </body>
 477 </html>