pserver_client.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>Design Doc: The Client Library of Parameter Server &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>
var _hmt = _hmt || [];
(function() {
  var hm = document.createElement("script");
  hm.src = "//hm.baidu.com/hm.js?b9a314ab40d04d805655aab1deee08ba";
  var s = document.getElementsByTagName("script")[0]; 
  s.parentNode.insertBefore(hm, s);
})();
</script>


  <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">
          
<nav class="doc-menu-vertical" role="navigation">

<ul>
<li class="toctree-l1"><a class="reference internal" href="../../getstarted/index_en.html">GET STARTED</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../getstarted/quickstart_en.html">Quick Start</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../getstarted/concepts/use_concepts_en.html">Basic Concept</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../build_and_install/index_en.html">Install and Build</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/pip_install_en.html">Install using pip</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/docker_install_en.html">Run in Docker Containers</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../build_and_install/build_from_source_en.html">Build from Sources</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../howto/index_en.html">HOW TO</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../howto/cmd_parameter/index_en.html">Set Command-line Parameters</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cmd_parameter/use_case_en.html">Use Case</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cmd_parameter/arguments_en.html">Argument Outline</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cmd_parameter/detail_introduction_en.html">Detail Description</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../../howto/cluster/index_en.html">Distributed Training</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cluster/preparations_en.html">Preparations</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cluster/cmd_argument_en.html">Command-line arguments</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/cluster/multi_cluster/index_en.html">Use different clusters</a><ul>
<li class="toctree-l4"><a class="reference internal" href="../../howto/cluster/multi_cluster/fabric_en.html">Fabric</a></li>
<li class="toctree-l4"><a class="reference internal" href="../../howto/cluster/multi_cluster/openmpi_en.html">OpenMPI</a></li>
<li class="toctree-l4"><a class="reference internal" href="../../howto/cluster/multi_cluster/k8s_en.html">Kubernetes</a></li>
<li class="toctree-l4"><a class="reference internal" href="../../howto/cluster/multi_cluster/k8s_aws_en.html">Kubernetes on AWS</a></li>
</ul>
</li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../../howto/rnn/index_en.html">RNN Models</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../howto/rnn/rnn_config_en.html">RNN Configuration</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/rnn/recurrent_group_en.html">Recurrent Group Tutorial</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/rnn/hierarchical_layer_en.html">Layers supporting hierarchical sequence as input</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../howto/rnn/hrnn_rnn_api_compare_en.html">API comparision between RNN and hierarchical RNN</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../../howto/optimization/gpu_profiling_en.html">Tune GPU Performance</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../dev/index_en.html">Development</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../dev/contribute_to_paddle_en.html">Contribute Code</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../dev/write_docs_en.html">Contribute Documentation</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../../faq/index_en.html">FAQ</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../faq/build_and_install/index_en.html">Install, Build and Unit test</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/model/index_en.html">Model Configuration</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/parameter/index_en.html">Parameter Setting</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/local/index_en.html">Local Training and Prediction</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../faq/cluster/index_en.html">Cluster Training and Prediction</a></li>
</ul>
</li>
</ul>

</nav>

        </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>Design Doc: The Client Library of Parameter Server</li>
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../../_sources/design/cluster_train/pserver_client.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="design-doc-the-client-library-of-parameter-server">
<span id="design-doc-the-client-library-of-parameter-server"></span><h1>Design Doc: The Client Library of Parameter Server<a class="headerlink" href="#design-doc-the-client-library-of-parameter-server" title="Permalink to this headline">¶</a></h1>
<p>For an overview of trainer&#8217;s role, please refer to <a class="reference internal" href="README.html"><span class="doc">distributed training design doc</span></a>. In this design doc, we will discuss the parameter server&#8217;s client library, which will manage communication with parameter servers. The library will be implemented in <a class="reference external" href="https://golang.org/">Go</a> and made available as a static or dynamic library with a C header file.</p>
<div class="section" id="parameter-partition">
<span id="parameter-partition"></span><h2>Parameter Partition<a class="headerlink" href="#parameter-partition" title="Permalink to this headline">¶</a></h2>
<p>Each parameter will be partitioned into parameter blocks to make the parameters evenly distributed on parameter servers. The partition is done automatically by the client library. The <em>sparse parameter</em> require a little different treatment:</p>
<div class="section" id="sparse-parameter">
<span id="sparse-parameter"></span><h3>Sparse Parameter<a class="headerlink" href="#sparse-parameter" title="Permalink to this headline">¶</a></h3>
<p>The sparse parameter is a parameter that is updated sparsely. The name is somewhat misleading, it does not have a sparse representation, it has the same representation as a dense vector.</p>
<p>Because a sparse parameter is updated sparsely, the trainer will have to partition the sparse parameter. Because the parameter server will merge all sparse parameter shard into the same file when saving the parameter. It needs special naming convention:</p>
<p>If a sparse parameter is partitioned into n shards, they should be named as:</p>
<div class="highlight-text"><div class="highlight"><pre><span></span>name:sparse-0
name:sparse-1
...
name:sparse-n-1
</pre></div>
</div>
<p>The library is unaware of the partition, and treat each parameter independently. Only when saving parameters, the parameter servers will merge the sparse parameters according to the naming convention.</p>
</div>
</div>
<div class="section" id="model-optimization-using-gradients">
<span id="model-optimization-using-gradients"></span><h2>Model Optimization Using Gradients<a class="headerlink" href="#model-optimization-using-gradients" title="Permalink to this headline">¶</a></h2>
<p>There are two ways to perform model optimization using gradients:</p>
<ul>
<li><p class="first">On Client</p>
<p>The client does multiple steps of forward and backward update. In each step, the gradients are calculated and a new model is generated. After some steps, the client will calculate the difference between the newest model and the old model at step 0. The difference will be updated to parameter servers. Parameter servers will just update parameters using the difference without any optimization using gradients (such as Adam and L1 regularization).</p>
</li>
<li><p class="first">On Parameter Server</p>
<p>The client will send accumulated gradients to parameter servers, the parameter server will do the optimization using gradients.</p>
</li>
</ul>
</div>
<div class="section" id="l1-and-l2-regularization">
<span id="l1-and-l2-regularization"></span><h2>L1 and L2 Regularization<a class="headerlink" href="#l1-and-l2-regularization" title="Permalink to this headline">¶</a></h2>
<p>PaddlePaddle allows L1 or L2 regularizations to be specified per parameter, so when the trainer initializes the parameter it needs include a parameter configuration when L1 or L2 regularization is necessary.</p>
</div>
<div class="section" id="parameter-initialization">
<span id="parameter-initialization"></span><h2>Parameter Initialization<a class="headerlink" href="#parameter-initialization" title="Permalink to this headline">¶</a></h2>
<p>The parameters on parameter servers need to be initialized. To provide maximum flexibility, the trainer will initialize the parameters. Only one trainer will do the initialization, the other trainers will wait for the completion of initialization and get the parameters from the parameter servers.</p>
<div class="section" id="trainer-selection">
<span id="trainer-selection"></span><h3>Trainer Selection<a class="headerlink" href="#trainer-selection" title="Permalink to this headline">¶</a></h3>
<p>To select the trainer for initialization, every trainer will try to get a distributed lock, whoever owns the lock will do the initialization. As illustrated below:</p>
<p><img src="./src/init_lock.png"></p>
</div>
<div class="section" id="trainer-selection-process">
<span id="trainer-selection-process"></span><h3>Trainer Selection Process<a class="headerlink" href="#trainer-selection-process" title="Permalink to this headline">¶</a></h3>
<p>The trainer select process is encapsulated in the C API function:</p>
<div class="highlight-c"><div class="highlight"><pre><span></span><span class="kt">int</span> <span class="nf">paddle_begin_init_params</span><span class="p">(</span><span class="n">paddle_pserver_client</span><span class="o">*</span> <span class="n">client</span><span class="p">,</span> <span class="k">const</span> <span class="kt">char</span><span class="o">*</span> <span class="n">config_proto</span><span class="p">);</span>
</pre></div>
</div>
<p>The selected trainer&#8217;s call to <code class="docutils literal"><span class="pre">paddle_begin_init_params</span></code> will return with 1, and the other trainers&#8217; call to <code class="docutils literal"><span class="pre">paddle_begin_init_params</span></code> will return 0. <code class="docutils literal"><span class="pre">paddle_get_params</span></code> will be blocked until initialization is completed. As illustrated below:</p>
<p><img src="./src/pserver_init.png"></p>
</div>
</div>
<div class="section" id="c-interface">
<span id="c-interface"></span><h2>C Interface<a class="headerlink" href="#c-interface" title="Permalink to this headline">¶</a></h2>
<div class="highlight-c"><div class="highlight"><pre><span></span><span class="k">typedef</span> <span class="k">enum</span> <span class="p">{</span>
  <span class="n">PADDLE_ELEMENT_TYPE_INT32</span>   <span class="o">=</span> <span class="mi">0</span><span class="p">,</span>
  <span class="n">PADDLE_ELEMENT_TYPE_UINT32</span>  <span class="o">=</span> <span class="mi">1</span><span class="p">,</span>
  <span class="n">PADDLE_ELEMENT_TYPE_INT64</span>   <span class="o">=</span> <span class="mi">2</span><span class="p">,</span>
  <span class="n">PADDLE_ELEMENT_TYPE_UINT64</span>  <span class="o">=</span> <span class="mi">3</span><span class="p">,</span>
  <span class="n">PADDLE_ELEMENT_TYPE_FLOAT32</span> <span class="o">=</span> <span class="mi">4</span><span class="p">,</span>
  <span class="n">PADDLE_ELEMENT_TYPE_FLOAT64</span> <span class="o">=</span> <span class="mi">5</span><span class="p">,</span>
<span class="p">}</span> <span class="n">paddle_element_type</span><span class="p">;</span>

<span class="k">typedef</span> <span class="k">struct</span> <span class="p">{</span>
  <span class="kt">char</span><span class="o">*</span>               <span class="n">name</span><span class="p">;</span>
  <span class="n">paddle_element_type</span> <span class="n">element_type</span><span class="p">;</span>
  <span class="kt">unsigned</span> <span class="kt">char</span><span class="o">*</span>      <span class="n">content</span><span class="p">;</span>
  <span class="kt">int</span>                 <span class="n">content_len</span><span class="p">;</span>
<span class="p">}</span> <span class="n">paddle_parameter</span><span class="p">,</span> <span class="n">paddle_gradient</span><span class="p">;</span>

<span class="k">typedef</span> <span class="kt">int</span> <span class="n">paddle_pserver_client</span><span class="p">;</span>

<span class="cm">/**</span>
<span class="cm"> * @brief creates a pserver client that talks to etcd for coordination.</span>
<span class="cm"> */</span>
<span class="n">paddle_pserver_client</span> <span class="nf">paddle_new_etcd_pserver_client</span><span class="p">(</span><span class="kt">char</span><span class="o">*</span> <span class="n">etcd_addr</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief creates a pserver client given pserver addresses.</span>
<span class="cm"> *</span>
<span class="cm"> * @param pserver_addrs comma-separated pserver addresses.</span>
<span class="cm"> * @param selected if current pserver client is selected to initialize all parameter servers.</span>
<span class="cm"> */</span>
<span class="n">paddle_pserver_client</span> <span class="nf">paddle_new_pserver_client</span><span class="p">(</span><span class="kt">char</span><span class="o">*</span> <span class="n">pserver_addrs</span><span class="p">,</span> <span class="kt">int</span> <span class="n">selected</span><span class="p">);</span>
<span class="kt">void</span> <span class="nf">paddle_pserver_client_release</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">c</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_begin_init_params begins to initialize parameters on</span>
<span class="cm"> * parameter servers.</span>
<span class="cm"> *</span>
<span class="cm"> * paddle_begin_init_params will be called from multiple trainers,</span>
<span class="cm"> * only one trainer will be selected to initialize the parameters on</span>
<span class="cm"> * parameter servers. Other trainers need to get the initialized</span>
<span class="cm"> * parameters from parameter servers using @paddle_get_params.</span>
<span class="cm"> *</span>
<span class="cm"> * @return 1 if the trainer is selected to initialize parameter</span>
<span class="cm"> * servers, otherwise 0.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_begin_init_params</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_init_param initializes the parameter on parameter</span>
<span class="cm"> * servers.</span>
<span class="cm"> *</span>
<span class="cm"> * @param param the parameter to initialize.</span>
<span class="cm"> * @param param_config_proto the configuration for the parameter.</span>
<span class="cm"> * @param config_len the length of param_config_proto</span>
<span class="cm"> * @return 0 if successful, otherwise -1. On failure, the trainer</span>
<span class="cm"> * needs to restart the entire initialization process (starting from</span>
<span class="cm"> * @paddle_begin_init_param). Or simply exit the program and wait for</span>
<span class="cm"> * the cluster management system to restart the trainer.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_init_param</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">,</span> <span class="n">paddle_parameter</span> <span class="n">param</span><span class="p">,</span> <span class="k">const</span> <span class="kt">unsigned</span> <span class="kt">char</span><span class="o">*</span> <span class="n">param_config_proto</span><span class="p">,</span> <span class="kt">int</span> <span class="n">config_len</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_finish_init_params tells parameter servers client has</span>
<span class="cm"> * sent all parameters to parameter servers as initialization.</span>
<span class="cm"> *</span>
<span class="cm"> * @return 0 if successful, otherwise -1. On failure, the trainer</span>
<span class="cm"> * needs to restart the entire initialization process (starting from</span>
<span class="cm"> * @paddle_begin_init_param). Or simply exit the program and wait for</span>
<span class="cm"> * the cluster management system to restart the trainer.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_finish_init_params</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_send_grads sends gradients to parameter servers for</span>
<span class="cm"> * updating parameters.</span>
<span class="cm"> *</span>
<span class="cm"> * @param grads the array of gradients to send.</span>
<span class="cm"> * @param len the length of the gradient array.</span>
<span class="cm"> * @param learning_rate the learning rate for the gradients.</span>
<span class="cm"> * @return 0 if successful, otherwise -1.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_send_grads</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">,</span> <span class="k">const</span> <span class="n">paddle_gradient</span><span class="o">*</span> <span class="n">grads</span><span class="p">,</span> <span class="kt">int</span> <span class="n">len</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_get_params gets parameters from parameter servers.</span>
<span class="cm"> *</span>
<span class="cm"> * paddle_get_params will block until parameters are initialized on</span>
<span class="cm"> * the parameter servers.</span>
<span class="cm"> *</span>
<span class="cm"> * @param dst the destination array of parameter pointers to save to.</span>
<span class="cm"> * The parameter pointer must be pre-popullated with required parameter name,</span>
<span class="cm"> * and the content of parameter must be pre-allocated of the size of required</span>
<span class="cm"> * parameter on pserver.</span>
<span class="cm"> * @param len the length of the names array and the paddle_parameter</span>
<span class="cm"> * array.</span>
<span class="cm"> * @return 0 if successful, otherwise -1.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_get_params</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">,</span> <span class="n">paddle_parameter</span><span class="o">**</span> <span class="n">dst</span><span class="p">,</span> <span class="kt">int</span> <span class="n">len</span><span class="p">);</span>

<span class="cm">/**</span>
<span class="cm"> * @brief paddle_save_model indicates parameters to save the parameter</span>
<span class="cm"> * to the given path</span>
<span class="cm"> *</span>
<span class="cm"> * @param path the path to save parameters.</span>
<span class="cm"> * @return 0 if successful, otherwise -1.</span>
<span class="cm"> */</span>
<span class="kt">int</span> <span class="nf">paddle_save_model</span><span class="p">(</span><span class="n">paddle_pserver_client</span> <span class="n">client</span><span class="p">,</span> <span class="k">const</span> <span class="kt">char</span><span class="o">*</span> <span class="n">path</span><span class="p">);</span>
</pre></div>
</div>
</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>