checkpointing.html 14.4 KB
Newer Older
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
<!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>模型参数检查点(Checkpointing) &mdash; PaddlePaddle  documentation</title>
  

  
  

  

  
  
    

  

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

  
33

34 35 36 37 38
  
        <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"/> 
39 40 41 42 43 44 45 46 47 48
<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>

49 50 51 52 53 54 55 56

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

</head>

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

57 58 59 60 61 62 63 64 65 66 67 68 69
  <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
          

          
70 71
          </a>

72 73 74 75 76 77
          
            
            
          

          
78 79 80 81 82 83
<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>
84
</div>
85 86

          
87 88 89 90
        </div>

        <div class="wy-menu wy-menu-vertical" data-spy="affix" role="navigation" aria-label="main navigation">
          
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
<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>
147 148
</ul>

149 150
</nav>

151 152
        </div>
      </div>
153 154
    </nav>

155
    <section data-toggle="wy-nav-shift" class="wy-nav-content-wrap">
156

157 158 159 160 161
      
      <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>
162 163


164 165 166 167
      
      <div class="wy-nav-content">
        <div class="rst-content">
          
168

169
 
170 171 172 173 174



<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
175
    <li><a href="../../index_en.html">Docs</a> &raquo;</li>
176 177
      
    <li>模型参数检查点(Checkpointing)</li>
178 179 180 181 182 183 184
      <li class="wy-breadcrumbs-aside">
        
          
            <a href="../../_sources/design/cluster_train/checkpointing.md.txt" rel="nofollow"> View page source</a>
          
        
      </li>
185
  </ul>
186
  <hr/>
187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279
</div>
          <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
           <div itemprop="articleBody">
            
  <div class="section" id="checkpointing">
<span id="checkpointing"></span><h1>模型参数检查点(Checkpointing)<a class="headerlink" href="#checkpointing" title="Permalink to this headline"></a></h1>
<p>模型数据检查点的实现,可以有效的避免parameter server的单点或多点同时故障。模型参数检查点通过定期向磁盘上保存一份存储在parameter server内存中的模型数据的完整镜像,来保证训练过程可以从中间状态重新启动。在一个不可中断并缺少备份的训练任务中,可以通过阶段性的保存每个parameter server的数据快照(snapshot)到 <strong><em>分布式存储服务</em></strong> 达到容灾的目的,比如每隔10分钟最新的快照,并删除更早的快照。在出现单点故障时,只需要恢复这台节点,或者将这台节点迁移到另一个节点并启动即可恢复训练任务。</p>
<p><img src="src/checkpointing.png" width="500"/></p>
<div class="section" id="">
<span id="id1"></span><h2>快照保存的设计如下:<a class="headerlink" href="#" title="Permalink to this headline"></a></h2>
<p>说明:</p>
<ul class="simple">
<li>parameter server在集群中启动后,自动挂载分布式存储目录,并把快照保存到这个目录下。</li>
<li><strong><em>注:每个parameter server的检查点各自独立保存,暂时不考虑多个parameter server同步的保存一个特定时间点的全局检查点,因为这样做也没法保证消除随机性。</em></strong></li>
</ul>
<p>检查点保存程序流程:</p>
<ol class="simple">
<li>如果满足条件&#8221;每隔10分钟&#8221;时,parameter server会获取parameters内存的<code class="docutils literal"><span class="pre">read_lock</span></code>,启动一个新的线程开始保存检查点。如果已经正在执行保存检查点的线程,则忽略。由于对parameters的更新需要获取parameters内存的<code class="docutils literal"><span class="pre">write_lock</span></code>,所以在写入快照的过程中,parameter server会暂停参数更新并等待。</li>
<li>parameter server生成一个UUID,向指定的目录中一个新的文件(文件名为此UUID)写入快照数据。在快照写入完成后,计算这个文件的MD5 sum。然后在etcd的<code class="docutils literal"><span class="pre">/checkpoints/[pserver_id]</span></code>中写入json内容:<code class="docutils literal"><span class="pre">{&quot;uuid&quot;:</span> <span class="pre">[UUID],</span> <span class="pre">&quot;md5&quot;,</span> <span class="pre">&quot;MD5</span> <span class="pre">sum&quot;,</span> <span class="pre">&quot;timestamp&quot;:</span> <span class="pre">xxxx}</span></code></li>
<li>删除磁盘目录中不是当前uuid的快照文件。</li>
<li>释放对paramters内存的锁定,停止保存检查点的线程。</li>
</ol>
<p>这里需要用户额外注意,在您的实际环境中,训练任务的运行可能会占满trainer和parameter server之间的网络带宽,如果parameter server此时还需要通过网络访问分布式存储以保存快照,可能会造成网络拥塞,而出现阶段性的运行停滞。</p>
</div>
<div class="section" id="">
<span id="id2"></span><h2>从快照恢复<a class="headerlink" href="#" title="Permalink to this headline"></a></h2>
<p>在parameter server第一次启动或任意时间parameter server故障后被Kubernetes重新启动,则需要回滚到上一个检查点:</p>
<ol class="simple">
<li>从etcd中读取节点:<code class="docutils literal"><span class="pre">/checkpoints/[pserver_id]</span></code>获取最新的检查点的文件uuid</li>
<li>从磁盘文件中加载uuid文件名的检查点快照文件,并加载其中的参数</li>
<li>如果上面两步出现错误,则使用启动参数定义的初始化方法初始化参数</li>
<li>开始提供服务</li>
</ol>
</div>
</div>
<div class="section" id="todo-list">
<span id="todo-list"></span><h1>TODO List<a class="headerlink" href="#todo-list" title="Permalink to this headline"></a></h1>
<div class="section" id="todo">
<span id="todo"></span><h2>推测执行/加速执行(TODO)<a class="headerlink" href="#todo" title="Permalink to this headline"></a></h2>
<p>在异构集群中,如果存在某些trainer执行速度过慢会影响整体集群的速度(如图中Trainer 1),此时master将负责启动一个新的Trainer(Accelerate Trainer 2),使用同样的训练数据block。哪个trainer先完成block的训练,则把另一个慢速的kill掉。</p>
</div>
<div class="section" id="">
<span id="id3"></span><h2>动态扩容/缩容<a class="headerlink" href="#" title="Permalink to this headline"></a></h2>
<p>目前只考虑动态扩容trainer数量,可以减小系统复杂性。</p>
</div>
</div>
<div class="section" id="">
<span id="id4"></span><h1>术语<a class="headerlink" href="#" title="Permalink to this headline"></a></h1>
<ul class="simple">
<li>model: 指深度学习训练之后得到的所有参数,使用这个神经网络可以完成对新数据的预测</li>
<li>parameters: 神经网络中的参数,包括权重w和偏置b。一个神经网络的模型由大量的参数组成</li>
<li>shard: 分片,通常指将一个整体拆分成多份的其中的一份。</li>
<li>model shard: 将一个神经网络参数拆分成多份,每个shard分别存储在其中一台parameter server之上</li>
<li>parameter block: 多个parameter block构成一个model shard</li>
<li>单点故障: 任意时刻只可能同时有一台服务器故障。由于集群中同时存在两台机器故障的概率极低((平均故障率*平均故障修复时间)^2)只对特殊在线系统考虑两台以上同时故障的容灾。</li>
</ul>
</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',
280
            HAS_SOURCE:  true
281 282 283 284 285 286
        };
    </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>
287

288 289 290 291 292 293
  

  
  
    <script type="text/javascript" src="../../_static/js/theme.js"></script>
  
294 295

  
296
  
297 298 299 300 301 302
  <script type="text/javascript">
      jQuery(function () {
          SphinxRtdTheme.StickyNav.enable();
      });
  </script>
   
303 304 305

</body>
</html>