large_model_dist_train.html



<!DOCTYPE html>
<!--[if IE 8]><html class="no-js lt-ie9" lang="en" > <![endif]-->
<!--[if gt IE 8]><!--> <html class="no-js" lang="en" > <!--<![endif]-->
<head>
  <meta charset="utf-8">
  
  <meta name="viewport" content="width=device-width, initial-scale=1.0">
  
  <title>Alalysis of large model distributed training in Paddle &mdash; PaddlePaddle  documentation</title>
  

    <link rel="stylesheet" href="../../_static/css/theme.css" type="text/css" />
  

        <link rel="index" title="Index"
              href="../../genindex.html"/>
        <link rel="search" title="Search" href="../../search.html"/>
    <link rel="top" title="PaddlePaddle  documentation" href="../../index.html"/> 

  
  <script src="../../_static/js/modernizr.min.js"></script>

</head>

<body class="wy-body-for-nav" role="document">

  <div class="wy-grid-for-nav">

    
    <nav data-toggle="wy-nav-shift" class="wy-nav-side">
      <div class="wy-side-scroll">
        <div class="wy-side-nav-search">
          

            <a href="../../index_en.html" class="icon icon-home"> PaddlePaddle
          

          </a>

          
<div role="search">
  <form id="rtd-search-form" class="wy-form" action="../../search.html" method="get">
    <input type="text" name="q" placeholder="Search docs" />
    <input type="hidden" name="check_keywords" value="yes" />
    <input type="hidden" name="area" value="default" />
  </form>
</div>

          
        </div>

        <div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
          
            
                <ul>
<li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../build_and_install/index_en.html">Install and Build</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../howto/index_en.html">HOW TO</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../dev/index_en.html">Development</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../faq/index_en.html">FAQ</a></li>
</ul>

            
        </div>
      </div>
    </nav>

    <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">

      
      <nav class="wy-nav-top" role="navigation" aria-label="top navigation">
        <i data-toggle="wy-nav-top" class="fa fa-bars"></i>
        <a href="../../index_en.html">PaddlePaddle</a>
      </nav>


      <div class="wy-nav-content">
        <div class="rst-content">
          

<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
    <li><a href="../../index_en.html">Docs</a> &raquo;</li>
      
    <li>Alalysis of large model distributed training in Paddle</li>
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../../_sources/design/cluster_train/large_model_dist_train.md.txt" rel="nofollow"> View page source</a>
          
        
      </li>
  </ul>
  <hr/>
</div>
          <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
           <div itemprop="articleBody">
            
  <div class="section" id="alalysis-of-large-model-distributed-training-in-paddle">
<span id="alalysis-of-large-model-distributed-training-in-paddle"></span><h1>Alalysis of large model distributed training in Paddle<a class="headerlink" href="#alalysis-of-large-model-distributed-training-in-paddle" title="Permalink to this headline">¶</a></h1>
<p><strong><em>NOTE: This is only some note for how we implemeted this scheme in V1, not a new design.</em></strong></p>
<div class="section" id="what-is-it">
<span id="what-is-it"></span><h2>What is it<a class="headerlink" href="#what-is-it" title="Permalink to this headline">¶</a></h2>
<p>We often encounter cases that the embedding layer parameters(sparse) are so large that we can not store it in the trainer&#8217;s memory when training. So we need to put them to several servers, and fetch them row by row instead of fetch all of the parameters.</p>
</div>
<div class="section" id="how-to-use">
<span id="how-to-use"></span><h2>How to use<a class="headerlink" href="#how-to-use" title="Permalink to this headline">¶</a></h2>
<p>Specify command-line argument like  <code class="docutils literal"><span class="pre">--loadsave_parameters_in_pserver=true</span> <span class="pre">--ports_num_for_sparse=1</span> <span class="pre">--use_old_updater=1</span></code> when starting the paddle trainer. And also add something like <code class="docutils literal"><span class="pre">--ports_num_for_sparse=1</span> <span class="pre">--pserver_num_threads=5</span></code> when starting pserver processes.</p>
<p>Accrodingly, configure your embedding layers like:</p>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="n">SPARSE_REMOTE</span><span class="o">=</span><span class="bp">True</span>

<span class="n">w1</span> <span class="o">=</span> <span class="n">data_layer</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s2">&quot;w1&quot;</span><span class="p">,</span> <span class="n">size</span><span class="o">=</span><span class="n">dict_size</span><span class="p">)</span>
<span class="n">emb1</span> <span class="o">=</span> <span class="n">embedding_layer</span><span class="p">(</span><span class="nb">input</span><span class="o">=</span><span class="n">w1</span><span class="p">,</span> <span class="n">size</span><span class="o">=</span><span class="mi">32</span><span class="p">,</span> <span class="n">param_attr</span><span class="o">=</span><span class="n">ParameterAttribute</span><span class="p">(</span><span class="n">sparse_update</span><span class="o">=</span><span class="n">SPARSE_REMOTE</span><span class="p">))</span>
<span class="n">w2</span> <span class="o">=</span> <span class="n">data_layer</span><span class="p">(</span><span class="n">name</span><span class="o">=</span><span class="s2">&quot;w2&quot;</span><span class="p">,</span> <span class="n">size</span><span class="o">=</span><span class="n">dict_size</span><span class="p">)</span>
<span class="n">emb2</span> <span class="o">=</span> <span class="n">embedding_layer</span><span class="p">(</span><span class="nb">input</span><span class="o">=</span><span class="n">w2</span><span class="p">,</span> <span class="n">size</span><span class="o">=</span><span class="mi">32</span><span class="p">,</span> <span class="n">param_attr</span><span class="o">=</span><span class="n">ParameterAttribute</span><span class="p">(</span><span class="n">sparse_update</span><span class="o">=</span><span class="n">SPARSE_REMOTE</span><span class="p">))</span>
<span class="o">...</span>
</pre></div>
</div>
</div>
<div class="section" id="implementation-details">
<span id="implementation-details"></span><h2>Implementation details<a class="headerlink" href="#implementation-details" title="Permalink to this headline">¶</a></h2>
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="k">enum</span> <span class="n">MatType</span> <span class="p">{</span>
  <span class="n">MAT_NORMAL</span><span class="p">,</span>
  <span class="n">MAT_NORMAL_SHARED</span><span class="p">,</span>
  <span class="n">MAT_VALUE_SHARED</span><span class="p">,</span>
  <span class="n">MAT_SPARSE_ROW_IDS</span><span class="p">,</span>
  <span class="n">MAT_SPARSE_ROW_AUTO_GROW</span><span class="p">,</span>
  <span class="n">MAT_CACHE_ROW</span><span class="p">,</span>
  <span class="n">MAT_SPARSE_ROW</span><span class="p">,</span>
  <span class="n">MAT_SPARSE_ROW_PREFETCH</span><span class="p">,</span>
  <span class="n">MAT_SPARSE_ROW_PREFETCH_FULL_SIZE</span><span class="p">,</span>
<span class="p">};</span>
</pre></div>
</div>
<p><code class="docutils literal"><span class="pre">MAT_SPARSE_ROW_PREFETCH</span></code> is what we use when configured to fetch only row of matrix when training.</p>
<p>In <code class="docutils literal"><span class="pre">trainer_internal.cpp:L93</span> <span class="pre">trainOneBatch</span></code>:</p>
<div class="highlight-c++"><div class="highlight"><pre><span></span>  <span class="k">if</span> <span class="p">(</span><span class="n">config_</span><span class="o">-&gt;</span><span class="n">getOptConfig</span><span class="p">().</span><span class="n">use_sparse_remote_updater</span><span class="p">())</span> <span class="p">{</span>
    <span class="n">REGISTER_TIMER</span><span class="p">(</span><span class="s">&quot;prefetch&quot;</span><span class="p">);</span>
    <span class="n">gradientMachine_</span><span class="o">-&gt;</span><span class="n">prefetch</span><span class="p">(</span><span class="n">inArgs</span><span class="p">);</span>
    <span class="n">parameterUpdater_</span><span class="o">-&gt;</span><span class="n">getParametersRemote</span><span class="p">();</span>
  <span class="p">}</span>
</pre></div>
</div>
<p>When doing actual network forward and backward, at the beginning of each batch, the trainer will try to download one row of data from pserver.</p>
<p>In <code class="docutils literal"><span class="pre">trainer/RemoteParameterUpdater.cpp</span></code>: <code class="docutils literal"><span class="pre">parameterUpdater_-&gt;getParametersRemote();</span></code>:</p>
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="k">if</span> <span class="p">(</span><span class="n">fullSize</span><span class="p">)</span> <span class="p">{</span>
    <span class="p">...</span>
<span class="p">}</span> <span class="k">else</span> <span class="p">{</span>
<span class="n">getParams</span> <span class="o">=</span> <span class="p">[</span><span class="o">&amp;</span><span class="p">]</span> <span class="p">{</span>
    <span class="n">parameterClient_</span><span class="o">-&gt;</span><span class="n">getParameterSparse</span><span class="p">(</span>
        <span class="cm">/* recvParameterType= */</span> <span class="n">PARAMETER_VALUE</span><span class="p">,</span> <span class="n">sendBackParameterType</span><span class="p">);</span>
<span class="p">};</span>
<span class="n">applyL1</span> <span class="o">=</span> <span class="p">[](</span><span class="n">Parameter</span><span class="o">&amp;</span> <span class="n">para</span><span class="p">,</span> <span class="n">real</span> <span class="n">decayRate</span><span class="p">)</span> <span class="p">{</span>
    <span class="n">para</span><span class="p">.</span><span class="n">getMat</span><span class="p">(</span><span class="n">PARAMETER_VALUE</span><span class="p">)</span><span class="o">-&gt;</span><span class="n">applyL1</span><span class="p">(</span><span class="cm">/*lr=*/</span><span class="mf">1.0f</span><span class="p">,</span> <span class="n">decayRate</span><span class="p">);</span>
<span class="p">};</span>
<span class="p">}</span>
</pre></div>
</div>
<p>Calling <code class="docutils literal"><span class="pre">parameterClient_-&gt;getParameterSparse</span></code> will do remote call to pserver&#8217;s <code class="docutils literal"><span class="pre">getParameterSparse</span></code>:</p>
<div class="highlight-c++"><div class="highlight"><pre><span></span><span class="kt">void</span> <span class="n">ParameterServer2</span><span class="o">::</span><span class="n">getParameterSparse</span><span class="p">(</span><span class="k">const</span> <span class="n">SendParameterRequest</span><span class="o">&amp;</span> <span class="n">request</span><span class="p">,</span>
                                          <span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o">&lt;</span><span class="n">Buffer</span><span class="o">&gt;&amp;</span> <span class="n">inputBuffers</span><span class="p">,</span>
                                          <span class="n">SendParameterResponse</span><span class="o">*</span> <span class="n">response</span><span class="p">,</span>
                                          <span class="n">std</span><span class="o">::</span><span class="n">vector</span><span class="o">&lt;</span><span class="n">Buffer</span><span class="o">&gt;*</span> <span class="n">outputBuffers</span><span class="p">)</span> <span class="p">{</span>
  <span class="p">(</span><span class="kt">void</span><span class="p">)</span><span class="n">inputBuffers</span><span class="p">;</span>
  <span class="k">auto</span><span class="o">&amp;</span> <span class="n">buffer</span> <span class="o">=</span> <span class="o">*</span><span class="n">readWriteBuffer_</span><span class="p">;</span>
  <span class="kt">size_t</span> <span class="n">numReals</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span>
  <span class="k">for</span> <span class="p">(</span><span class="k">const</span> <span class="k">auto</span><span class="o">&amp;</span> <span class="nl">block</span> <span class="p">:</span> <span class="n">request</span><span class="p">.</span><span class="n">blocks</span><span class="p">())</span> <span class="p">{</span>
    <span class="n">numReals</span> <span class="o">+=</span> <span class="n">getParameterConfig</span><span class="p">(</span><span class="n">block</span><span class="p">).</span><span class="n">dims</span><span class="p">(</span><span class="mi">1</span><span class="p">);</span>
  <span class="p">}</span>
  <span class="n">buffer</span><span class="p">.</span><span class="n">resize</span><span class="p">(</span><span class="n">numReals</span><span class="p">);</span>

  <span class="n">VLOG</span><span class="p">(</span><span class="mi">3</span><span class="p">)</span> <span class="o">&lt;&lt;</span> <span class="s">&quot;pserver: getParameterSparse, numReals=&quot;</span> <span class="o">&lt;&lt;</span> <span class="n">numReals</span><span class="p">;</span>

  <span class="n">ReadLockGuard</span> <span class="nf">guard</span><span class="p">(</span><span class="n">parameterMutex_</span><span class="p">);</span>
  <span class="kt">size_t</span> <span class="n">offset</span> <span class="o">=</span> <span class="mi">0</span><span class="p">;</span>
  <span class="k">for</span> <span class="p">(</span><span class="k">const</span> <span class="k">auto</span><span class="o">&amp;</span> <span class="nl">block</span> <span class="p">:</span> <span class="n">request</span><span class="p">.</span><span class="n">blocks</span><span class="p">())</span> <span class="p">{</span>
    <span class="kt">size_t</span> <span class="n">width</span> <span class="o">=</span> <span class="n">getParameterConfig</span><span class="p">(</span><span class="n">block</span><span class="p">).</span><span class="n">dims</span><span class="p">(</span><span class="mi">1</span><span class="p">);</span>
    <span class="n">Buffer</span> <span class="n">buf</span> <span class="o">=</span> <span class="p">{</span><span class="n">buffer</span><span class="p">.</span><span class="n">data</span><span class="p">()</span> <span class="o">+</span> <span class="n">offset</span><span class="p">,</span> <span class="n">width</span><span class="p">};</span>
    <span class="kt">int</span> <span class="n">type</span> <span class="o">=</span> <span class="n">request</span><span class="p">.</span><span class="n">send_back_parameter_type</span><span class="p">();</span>
    <span class="n">sendBackParameterSparse</span><span class="p">(</span><span class="n">block</span><span class="p">,</span> <span class="n">type</span><span class="p">,</span> <span class="n">response</span><span class="p">,</span> <span class="o">&amp;</span><span class="n">buf</span><span class="p">,</span> <span class="n">width</span><span class="p">,</span> <span class="n">outputBuffers</span><span class="p">);</span>
    <span class="n">offset</span> <span class="o">+=</span> <span class="n">width</span><span class="p">;</span>
  <span class="p">}</span>
<span class="p">}</span>
</pre></div>
</div>
<p><code class="docutils literal"><span class="pre">getParameterConfig(block).dims(1)</span></code> returns the width of the current &#8220;parameter block&#8221;(a shard of parameter object),
then <code class="docutils literal"><span class="pre">getParameterSparse</span></code> remote call returns only one row of data to the client.</p>
</div>
</div>


           </div>
          </div>
          <footer>
  

  <hr/>

  <div role="contentinfo">
    <p>
        &copy; Copyright 2016, PaddlePaddle developers.

    </p>
  </div>
  Built with <a href="http://sphinx-doc.org/">Sphinx</a> using a <a href="https://github.com/snide/sphinx_rtd_theme">theme</a> provided by <a href="https://readthedocs.org">Read the Docs</a>. 

</footer>

        </div>
      </div>

    </section>

  </div>
  

    <script type="text/javascript">
        var DOCUMENTATION_OPTIONS = {
            URL_ROOT:'../../',
            VERSION:'',
            COLLAPSE_INDEX:false,
            FILE_SUFFIX:'.html',
            HAS_SOURCE:  true
        };
    </script>
      <script type="text/javascript" src="../../_static/jquery.js"></script>
      <script type="text/javascript" src="../../_static/underscore.js"></script>
      <script type="text/javascript" src="../../_static/doctools.js"></script>
      <script type="text/javascript" src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.0/MathJax.js?config=TeX-AMS-MML_HTMLorMML"></script>

  
    <script type="text/javascript" src="../../_static/js/theme.js"></script>
  

  <script type="text/javascript">
      jQuery(function () {
          SphinxRtdTheme.StickyNav.enable();
      });
  </script>
   

</body>
</html>