Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2299
Star 20931
Fork 5422
P
Paddle
切换分支/标签
查找文件 普通视图历史永久链接Permalink
k8s_distributed_cn.html 52.3 KB
Newer Older
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 


<!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>Kubernetes分布式训练 &mdash; PaddlePaddle  文档</title>
  

  
  

  

  
  
    

  

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

  
  
        <link rel="index" title="索引"
              href="../../../genindex.html"/>
        <link rel="search" title="搜索" href="../../../search.html"/>
    <link rel="top" title="PaddlePaddle  文档" href="../../../index.html"/>
T
Deploy to GitHub Pages: 33e074dab85d23bccdb1876df6490a7a5514e395  
Travis CI 已提交
36 
        <link rel="up" title="分布式训练" href="cluster_train_cn.html"/>
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
37 
        <link rel="next" title="Distributed PaddlePaddle Training on AWS with Kubernetes" href="k8s_aws_cn.html"/>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 
        <link rel="prev" title="Kubernetes单机训练" href="k8s_cn.html"/> 

  <link rel="stylesheet" href="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/css/perfect-scrollbar.min.css" type="text/css" />
  <link rel="stylesheet" href="../../../_static/css/override.css" type="text/css" />
  <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">

  
  <header class="site-header">
    <div class="site-logo">
      <a href="/"><img src="../../../_static/images/PP_w.png"></a>
    </div>
    <div class="site-nav-links">
      <div class="site-menu">
T
Deploy to GitHub Pages: 7b81055395daa0ce3d02ccb060476c410ee4eb0f  
Travis CI 已提交
68 
        <a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Fork me on Github</a>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
69 70 71 72 73 74 75 76 77 78 79 80 
        <div class="language-switcher dropdown">
          <a type="button" data-toggle="dropdown">
            <span>English</span>
            <i class="fa fa-angle-up"></i>
            <i class="fa fa-angle-down"></i>
          </a>
          <ul class="dropdown-menu">
            <li><a href="/doc_cn">中文</a></li>
            <li><a href="/doc">English</a></li>
          </ul>
        </div>
        <ul class="site-page-links">
T
Deploy to GitHub Pages: 18e5edf6a5b09868dbb083c312fe852cb8c29e55  
Travis CI 已提交
81 
          <li><a href="/">Home</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
82 83 84 85 86 87 88 89 90 
        </ul>
      </div>
      <div class="doc-module">
        
        <ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../../../getstarted/index_cn.html">新手入门</a></li>
<li class="toctree-l1 current"><a class="reference internal" href="../../index_cn.html">进阶指南</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../api/index_cn.html">API</a></li>
<li class="toctree-l1"><a class="reference internal" href="../../../faq/index_cn.html">FAQ</a></li>
T
Deploy to GitHub Pages: 664691c89af182de3919fc72a7c875311e03893d  
Travis CI 已提交
91 
<li class="toctree-l1"><a class="reference internal" href="../../../mobile/index_cn.html">MOBILE</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 
</ul>

        
<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>
  </header>
  
  <div class="main-content-wrap">

    
    <nav class="doc-menu-vertical" role="navigation">
        
          
          <ul class="current">
<li class="toctree-l1"><a class="reference internal" href="../../../getstarted/index_cn.html">新手入门</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../getstarted/build_and_install/index_cn.html">安装与编译</a><ul>
T
Deploy to GitHub Pages: 0c0ff2828ccedb51db23290d6df9e4c83839d6af  
Travis CI 已提交
115 116 
<li class="toctree-l3"><a class="reference internal" href="../../../getstarted/build_and_install/pip_install_cn.html">使用pip安装</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../getstarted/build_and_install/docker_install_cn.html">使用Docker安装运行</a></li>
T
Deploy to GitHub Pages: 458ffbf4fe7fe6da5e2ea133c027d560815cc3cc  
Travis CI 已提交
117 
<li class="toctree-l3"><a class="reference internal" href="../../dev/build_cn.html">用Docker编译和测试PaddlePaddle</a></li>
T
Deploy to GitHub Pages: 0c0ff2828ccedb51db23290d6df9e4c83839d6af  
Travis CI 已提交
118 
<li class="toctree-l3"><a class="reference internal" href="../../../getstarted/build_and_install/build_from_source_cn.html">从源码编译</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
119 120 
</ul>
</li>
T
Deploy to GitHub Pages: ab5ad2215d037ed6985f0830669c4bde8944a944  
Travis CI 已提交
121 
<li class="toctree-l2"><a class="reference internal" href="../../../getstarted/concepts/use_concepts_cn.html">基本使用概念</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
122 123 124 125 126 127 128 129 130 
</ul>
</li>
<li class="toctree-l1 current"><a class="reference internal" href="../../index_cn.html">进阶指南</a><ul class="current">
<li class="toctree-l2"><a class="reference internal" href="../cmd_parameter/index_cn.html">设置命令行参数</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/use_case_cn.html">使用案例</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/arguments_cn.html">参数概述</a></li>
<li class="toctree-l3"><a class="reference internal" href="../cmd_parameter/detail_introduction_cn.html">细节描述</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: 33e074dab85d23bccdb1876df6490a7a5514e395  
Travis CI 已提交
131 132 133 134 135 136 
<li class="toctree-l2 current"><a class="reference internal" href="cluster_train_cn.html">分布式训练</a><ul class="current">
<li class="toctree-l3"><a class="reference internal" href="fabric_cn.html">fabric集群</a></li>
<li class="toctree-l3"><a class="reference internal" href="openmpi_cn.html">openmpi集群</a></li>
<li class="toctree-l3"><a class="reference internal" href="k8s_cn.html">kubernetes单机</a></li>
<li class="toctree-l3 current"><a class="current reference internal" href="#">kubernetes distributed分布式</a></li>
<li class="toctree-l3"><a class="reference internal" href="k8s_aws_cn.html">AWS上运行kubernetes集群训练</a></li>
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
137 138 
</ul>
</li>
T
Deploy to GitHub Pages: 70ee79c2567d5128a9d527b444fcee30f83c02a5  
Travis CI 已提交
139 
<li class="toctree-l2"><a class="reference internal" href="../../dev/contribute_to_paddle_cn.html">如何贡献代码</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
140 141 
<li class="toctree-l2"><a class="reference internal" href="../../dev/write_docs_cn.html">如何贡献/修改文档</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../deep_model/rnn/index_cn.html">RNN相关模型</a><ul>
T
Deploy to GitHub Pages: 7c088b529c97f6dc3c28d60dd0510a64bce61312  
Travis CI 已提交
142 
<li class="toctree-l3"><a class="reference internal" href="../../deep_model/rnn/rnn_config_cn.html">RNN配置</a></li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
143 144 145 146 147 148 149 150 
<li class="toctree-l3"><a class="reference internal" href="../../deep_model/rnn/recurrent_group_cn.html">Recurrent Group教程</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../deep_model/rnn/hierarchical_layer_cn.html">支持双层序列作为输入的Layer</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../deep_model/rnn/hrnn_rnn_api_compare_cn.html">单双层RNN API对比介绍</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../../optimization/gpu_profiling_cn.html">GPU性能分析与调优</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: 5f2cbce4190fb83883fed45668b5e4d87a577589  
Travis CI 已提交
151 
<li class="toctree-l1"><a class="reference internal" href="../../../api/index_cn.html">API</a><ul>
T
Deploy to GitHub Pages: 362a14d777ec795f1f6b397c1039ebb594050a10  
Travis CI 已提交
152 153 154 
<li class="toctree-l2"><a class="reference internal" href="../../../api/v2/model_configs.html">模型配置</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/activation.html">Activation</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/layer.html">Layers</a></li>
T
Deploy to GitHub Pages: 0beda8736ffcac80a95bea20f873d8972d2b5b5e  
Travis CI 已提交
155 
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/evaluators.html">Evaluators</a></li>
T
Deploy to GitHub Pages: 362a14d777ec795f1f6b397c1039ebb594050a10  
Travis CI 已提交
156 157 158 159 160 161 
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/optimizer.html">Optimizer</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/pooling.html">Pooling</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/networks.html">Networks</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/config/attr.html">Parameter Attribute</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: d4d8f742a3bc4366bf410bc9ff29c0c32666bf60  
Travis CI 已提交
162 163 164 165 166 167 
<li class="toctree-l2"><a class="reference internal" href="../../../api/v2/data.html">数据访问</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/data/data_reader.html">Data Reader Interface</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/data/image.html">Image Interface</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/data/dataset.html">Dataset</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: 362a14d777ec795f1f6b397c1039ebb594050a10  
Travis CI 已提交
168 
<li class="toctree-l2"><a class="reference internal" href="../../../api/v2/run_logic.html">训练与应用</a></li>
T
Deploy to GitHub Pages: 5ac72d95014e453c13a389248d2d2361a45a3402  
Travis CI 已提交
169 170 171 172 173 174 175 176 177 178 179 180 181 
<li class="toctree-l2"><a class="reference internal" href="../../../api/v2/fluid.html">Fluid</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/layers.html">Layers</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/data_feeder.html">DataFeeder</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/executor.html">Executor</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/initializer.html">Initializer</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/evaluator.html">Evaluator</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/nets.html">Nets</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/optimizer.html">Optimizer</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/param_attr.html">ParamAttr</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/profiler.html">Profiler</a></li>
<li class="toctree-l3"><a class="reference internal" href="../../../api/v2/fluid/regularizer.html">Regularizer</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: 5f2cbce4190fb83883fed45668b5e4d87a577589  
Travis CI 已提交
182 183 
</ul>
</li>
T
Deploy to GitHub Pages: a703a246428dd9f95775db102e338aebf93e9764  
Travis CI 已提交
184 185 186 187 188 189 190 191 
<li class="toctree-l1"><a class="reference internal" href="../../../faq/index_cn.html">FAQ</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../../../faq/build_and_install/index_cn.html">编译安装与单元测试</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../faq/model/index_cn.html">模型配置</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../faq/parameter/index_cn.html">参数设置</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../faq/local/index_cn.html">本地训练与预测</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../faq/cluster/index_cn.html">集群训练与预测</a></li>
</ul>
</li>
T
Deploy to GitHub Pages: 664691c89af182de3919fc72a7c875311e03893d  
Travis CI 已提交
192 
<li class="toctree-l1"><a class="reference internal" href="../../../mobile/index_cn.html">MOBILE</a><ul>
T
Deploy to GitHub Pages: c808fbbfcbaaf5c08f6254bfdb860f5dac76a627  
Travis CI 已提交
193 194 195 
<li class="toctree-l2"><a class="reference internal" href="../../../mobile/cross_compiling_for_android_cn.html">Android平台编译指南</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../mobile/cross_compiling_for_ios_cn.html">iOS平台编译指南</a></li>
<li class="toctree-l2"><a class="reference internal" href="../../../mobile/cross_compiling_for_raspberry_cn.html">Raspberry Pi平台编译指南</a></li>
T
Deploy to GitHub Pages: 664691c89af182de3919fc72a7c875311e03893d  
Travis CI 已提交
196 197 
</ul>
</li>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 
</ul>

        
    </nav>
    
    <section class="doc-content-wrap">

      

 







<div role="navigation" aria-label="breadcrumbs navigation">
  <ul class="wy-breadcrumbs">
      
        <li><a href="../../index_cn.html">进阶指南</a> > </li>
T
Deploy to GitHub Pages: 33e074dab85d23bccdb1876df6490a7a5514e395  
Travis CI 已提交
220 
        <li><a href="cluster_train_cn.html">分布式训练</a> > </li>
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
221
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
222 223 224 225 226 227 228 229 230 231 232 
    <li>Kubernetes分布式训练</li>
  </ul>
</div>
      
      <div class="wy-nav-content" id="doc-content">
        <div class="rst-content">
          <div role="main" class="document" itemscope="itemscope" itemtype="http://schema.org/Article">
           <div itemprop="articleBody">
            
  <div class="section" id="kubernetes">
<span id="kubernetes"></span><h1>Kubernetes分布式训练<a class="headerlink" href="#kubernetes" title="永久链接至标题"></a></h1>
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
233 
<p>前一篇文章介绍了如何在Kubernetes集群上启动一个单机PaddlePaddle训练作业 (Job)。在这篇文章里，我们介绍如何在Kubernetes集群上进行分布式PaddlePaddle训练作业。关于PaddlePaddle的分布式训练，文章 <a class="reference external" href="http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cluster/cluster_train_cn.html">Cluster Training</a>介绍了一种通过SSH远程分发任务，进行分布式训练的方法，与此不同的是，本文将介绍在Kubernetes容器管理平台上快速构建PaddlePaddle容器集群，进行分布式训练的方案。</p>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 
<div class="section" id="">
<span id="id1"></span><h2>整体方案<a class="headerlink" href="#" title="永久链接至标题"></a></h2>
<p>在训练之前，用户将配置与训练数据切分好放在分布式文件系统预先分配好的目录中(不同的分布式文件系统，需要使用其制定的方式挂载后并导入数据)，训练时，程序从此目录拷贝文件到容器内进行训练，将结果保存到此目录里。整体的结构图如下：</p>
<p><img alt="paddle on kubernetes结构图" src="../../../_images/k8s-paddle-arch.png" /></p>
<p>上图描述了一个3节点的分布式训练场景，在每个Pod上都通过volume方式挂载分布式文件系统的一个目录用于保存训练数据和输出结果。Kubernetes为这次训练创建了3个pod并且调度到了3个node上运行，每个pod包含一个PaddlePaddle容器。在容器创建后，会启动pserver与trainer进程，读取volume中的数据进行这次分布式训练。</p>
<p>根据前文的描述，要在已有的Kubernetes集群上进行PaddlePaddle的分布式训练，按照下面步骤即可：</p>
<ol class="simple">
<li><a class="reference external" href="#制作镜像">制作PaddlePaddle镜像</a></li>
<li><a class="reference external" href="#上传训练文件">将训练文件与切分好的数据上传到共享存储</a></li>
<li><a class="reference external" href="#创建Job">编写本次训练的YAML文件，创建一个Kubernetes job</a></li>
<li><a class="reference external" href="#查看输出">训练结束后查看输出结果</a></li>
</ol>
<p>下面就根据这几个步骤分别介绍。</p>
<div class="section" id="">
<span id="id2"></span><h3>制作镜像<a class="headerlink" href="#" title="永久链接至标题"></a></h3>
<p>PaddlePaddle镜像需要提供<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>与<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>进程的运行环境，用这个镜像创建的容器需要有以下两个功能：</p>
<ul class="simple">
<li>拷贝训练文件到容器内</li>
<li>生成<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>与<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>进程的启动参数，并且启动训练</li>
</ul>
T
Deploy to GitHub Pages: a87f4963edf35dcf4b203fb6e2541f04b424dc02  
Travis CI 已提交
254 
<p>因为官方镜像 <code class="docutils literal"><span class="pre">paddlepaddle/paddle:latest</span></code> 内已经包含PaddlePaddle的执行程序但是还没上述功能，所以我们可以在这个基础上，添加启动脚本，制作新镜像来完成以上的工作。参考镜像的<a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/doc/howto/usage/cluster/src/k8s_train/Dockerfile"><em>Dockerfile</em></a>。</p>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
255 256 257 258 259 260 261 262 263 264 265 
<div class="highlight-bash"><div class="highlight"><pre><span></span>$ <span class="nb">cd</span> doc/howto/usage/k8s/src/k8s_train
$ docker build -t <span class="o">[</span>YOUR_REPO<span class="o">]</span>/paddle:mypaddle .
</pre></div>
</div>
<p>然后将构建成功的镜像上传到镜像仓库。</p>
<div class="highlight-bash"><div class="highlight"><pre><span></span>docker push  <span class="o">[</span>YOUR_REPO<span class="o">]</span>/paddle:mypaddle
</pre></div>
</div>
<p>注意上述命令中<code class="docutils literal"><span class="pre">[YOUR_REPO]</span></code>表示读者所使用的Docker镜像仓库地址，读者需要替换成自己使用的仓库地址。下文使用<code class="docutils literal"><span class="pre">[YOUR_REPO]/paddle:mypaddle</span></code>这个地址来表示此步骤所构建出的镜像。</p>
</div>
<div class="section" id="">
T
Deploy to GitHub Pages: 8bb2613f9cf99f7a52b885244e162df37c00f7a2  
Travis CI 已提交
266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 
<span id="id3"></span><h3>准备训练数据<a class="headerlink" href="#" title="永久链接至标题"></a></h3>
<p>这里我们通过在Kubernetes集群上启动一个Job来下载并切割数据，也可以通过修改<a class="reference internal" href="src/k8s_train/README.html"><span class="doc">k8s</span></a>的内容来定制image.</p>
<p>在启动Job之前，需要根据不同的分布式存储来绑定一个<a class="reference external" href="https://kubernetes.io/docs/user-guide/persistent-volumes/">persistentVolumeClaim</a>,生成的数据将会存储在这个volume下.</p>
<div class="highlight-yaml"><div class="highlight"><pre><span></span><span class="l l-Scalar l-Scalar-Plain">apiVersion</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">batch/v1</span>
<span class="l l-Scalar l-Scalar-Plain">kind</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">Job</span>
<span class="l l-Scalar l-Scalar-Plain">metadata</span><span class="p p-Indicator">:</span>
  <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddle-data</span>
<span class="l l-Scalar l-Scalar-Plain">spec</span><span class="p p-Indicator">:</span>
  <span class="l l-Scalar l-Scalar-Plain">template</span><span class="p p-Indicator">:</span>
    <span class="l l-Scalar l-Scalar-Plain">metadata</span><span class="p p-Indicator">:</span>
      <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">pi</span>
    <span class="l l-Scalar l-Scalar-Plain">spec</span><span class="p p-Indicator">:</span>
      <span class="l l-Scalar l-Scalar-Plain">hostNetwork</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">true</span>
      <span class="l l-Scalar l-Scalar-Plain">containers</span><span class="p p-Indicator">:</span>
      <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddle-data</span>
T
Deploy to GitHub Pages: a87f4963edf35dcf4b203fb6e2541f04b424dc02  
Travis CI 已提交
281 
        <span class="l l-Scalar l-Scalar-Plain">image</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddlepaddle/paddle-tutorial:k8s_data</span>
T
Deploy to GitHub Pages: 8bb2613f9cf99f7a52b885244e162df37c00f7a2  
Travis CI 已提交
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 
        <span class="l l-Scalar l-Scalar-Plain">imagePullPolicy</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">Always</span>
        <span class="l l-Scalar l-Scalar-Plain">volumeMounts</span><span class="p p-Indicator">:</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">mountPath</span><span class="p p-Indicator">:</span> <span class="s">&quot;/mnt&quot;</span>
          <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">nfs</span>
        <span class="l l-Scalar l-Scalar-Plain">env</span><span class="p p-Indicator">:</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">OUT_DIR</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">/home/work/mfs/paddle-cluster-job</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">SPLIT_COUNT</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="s">&quot;3&quot;</span>
      <span class="l l-Scalar l-Scalar-Plain">volumes</span><span class="p p-Indicator">:</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">nfs</span>
          <span class="l l-Scalar l-Scalar-Plain">persistentVolumeClaim</span><span class="p p-Indicator">:</span>
            <span class="l l-Scalar l-Scalar-Plain">claimName</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">mfs</span>
      <span class="l l-Scalar l-Scalar-Plain">restartPolicy</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">Never</span>
</pre></div>
</div>
<p>完成后volume中的文件内容大致如下：</p>
<div class="highlight-base"><div class="highlight"><pre><span></span>[root@paddle-kubernetes-node0 nfsdir]$ tree -d
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
300 
.
T
Deploy to GitHub Pages: 8bb2613f9cf99f7a52b885244e162df37c00f7a2  
Travis CI 已提交
301 302 303 304 305 306 307 308 309 
`-- paddle-cluster-job
    |-- 0
    |   `-- data
    |-- 1
    |   `-- data
    |-- 2
    |   `-- data
    |-- output
    |-- quick_start
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 
</pre></div>
</div>
<p>目录中paddle-cluster-job是本次训练对应的job name，本次训练要求有3个PaddlePaddle节点，在paddle-cluster-job/data目录中存放切分好的数据，文件夹0，1，2分别代表3个节点的trainer_id。recommendation文件夹内存放训练文件，output文件夹存放训练结果与日志。</p>
</div>
<div class="section" id="job">
<span id="job"></span><h3>创建Job<a class="headerlink" href="#job" title="永久链接至标题"></a></h3>
<p>Kubernetes可以通过YAML文件来创建相关对象，然后可以使用命令行工具创建job。</p>
<p>Job YAML文件描述了这次训练使用的Docker镜像，需要启动的节点个数以及 <code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>与 <code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>进程启动的必要参数，也描述了容器需要使用的存储卷挂载的情况。YAML文件中各个字段的具体含义，可以查看<a class="reference external" href="http://kubernetes.io/docs/api-reference/batch/v1/definitions/#_v1_job">Kubernetes Job API</a>。例如，本次训练的YAML文件可以写成：</p>
<div class="highlight-yaml"><div class="highlight"><pre><span></span><span class="l l-Scalar l-Scalar-Plain">apiVersion</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">batch/v1</span>
<span class="l l-Scalar l-Scalar-Plain">kind</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">Job</span>
<span class="l l-Scalar l-Scalar-Plain">metadata</span><span class="p p-Indicator">:</span>
  <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddle-cluster-job</span>
<span class="l l-Scalar l-Scalar-Plain">spec</span><span class="p p-Indicator">:</span>
  <span class="l l-Scalar l-Scalar-Plain">parallelism</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">3</span>
  <span class="l l-Scalar l-Scalar-Plain">completions</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">3</span>
  <span class="l l-Scalar l-Scalar-Plain">template</span><span class="p p-Indicator">:</span>
    <span class="l l-Scalar l-Scalar-Plain">metadata</span><span class="p p-Indicator">:</span>
      <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddle-cluster-job</span>
    <span class="l l-Scalar l-Scalar-Plain">spec</span><span class="p p-Indicator">:</span>
      <span class="l l-Scalar l-Scalar-Plain">volumes</span><span class="p p-Indicator">:</span>
      <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">jobpath</span>
        <span class="l l-Scalar l-Scalar-Plain">hostPath</span><span class="p p-Indicator">:</span>
          <span class="l l-Scalar l-Scalar-Plain">path</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">/home/work/mfs</span>
      <span class="l l-Scalar l-Scalar-Plain">containers</span><span class="p p-Indicator">:</span>
      <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">trainer</span>
        <span class="l l-Scalar l-Scalar-Plain">image</span><span class="p p-Indicator">:</span> <span class="p p-Indicator">[</span><span class="nv">YOUR_REPO</span><span class="p p-Indicator">]</span><span class="l l-Scalar l-Scalar-Plain">/paddle:mypaddle</span>
        <span class="l l-Scalar l-Scalar-Plain">command</span><span class="p p-Indicator">:</span> <span class="p p-Indicator">[</span><span class="s">&quot;bin/bash&quot;</span><span class="p p-Indicator">,</span>  <span class="s">&quot;-c&quot;</span><span class="p p-Indicator">,</span> <span class="s">&quot;/root/start.sh&quot;</span><span class="p p-Indicator">]</span>
        <span class="l l-Scalar l-Scalar-Plain">env</span><span class="p p-Indicator">:</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">JOB_NAME</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">paddle-cluster-job</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">JOB_PATH</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">/home/jobpath</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">JOB_NAMESPACE</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">default</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">TRAIN_CONFIG_DIR</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">recommendation</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">CONF_PADDLE_NIC</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">eth0</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">CONF_PADDLE_PORT</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="s">&quot;7164&quot;</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">CONF_PADDLE_PORTS_NUM</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="s">&quot;2&quot;</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">CONF_PADDLE_PORTS_NUM_SPARSE</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="s">&quot;2&quot;</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">CONF_PADDLE_GRADIENT_NUM</span>
          <span class="l l-Scalar l-Scalar-Plain">value</span><span class="p p-Indicator">:</span> <span class="s">&quot;3&quot;</span>
        <span class="l l-Scalar l-Scalar-Plain">volumeMounts</span><span class="p p-Indicator">:</span>
        <span class="p p-Indicator">-</span> <span class="l l-Scalar l-Scalar-Plain">name</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">jobpath</span>
          <span class="l l-Scalar l-Scalar-Plain">mountPath</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">/home/jobpath</span>
      <span class="l l-Scalar l-Scalar-Plain">restartPolicy</span><span class="p p-Indicator">:</span> <span class="l l-Scalar l-Scalar-Plain">Never</span>
</pre></div>
</div>
<p>文件中，<code class="docutils literal"><span class="pre">metadata</span></code>下的<code class="docutils literal"><span class="pre">name</span></code>表示这个job的名字。<code class="docutils literal"><span class="pre">parallelism，completions</span></code>字段表示这个job会同时开启3个PaddlePaddle节点，成功训练且退出的pod数目为3时，这个job才算成功结束。然后申明一个存储卷<code class="docutils literal"><span class="pre">jobpath</span></code>，代表宿主机目录<code class="docutils literal"><span class="pre">/home/work/mfs</span></code>，在对容器的描述<code class="docutils literal"><span class="pre">containers</span></code>字段中，将此目录挂载为容器的<code class="docutils literal"><span class="pre">/home/jobpath</span></code>目录，这样容器的<code class="docutils literal"><span class="pre">/home/jobpath</span></code>目录就成为了共享存储，放在这个目录里的文件其实是保存到了MFS上。</p>
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
363 364 365 366 367 368 369 370 371 372 373 374 
<p><code class="docutils literal"><span class="pre">env</span></code>字段表示容器的环境变量，我们将<code class="docutils literal"><span class="pre">paddle</span></code>运行的一些参数通过这种方式传递到容器内：</p>
<ul class="simple">
<li>JOB_PATH：共享存储挂在的路径</li>
<li>JOB_NAME：Job的名字</li>
<li>TRAIN_CONFIG_DIR：本次训练文件所在目录，与JOB_PATH,JOB_NAME组合可以找到本次训练需要的文件路径</li>
<li>CONF_PADDLE_NIC：<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>进程需要的<code class="docutils literal"><span class="pre">--nics</span></code>参数，即网卡名</li>
<li>CONF_PADDLE_PORT：<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">paserver</span></code>的<code class="docutils literal"><span class="pre">--port</span></code>参数</li>
<li>CONF_PADDLE_PORTS_NUM：稠密更新的端口数量，即<code class="docutils literal"><span class="pre">--ports_num</span></code>参数</li>
<li>CONF_PADDLE_PORTS_NUM_SPARSE：稀疏更新的端口数量，即<code class="docutils literal"><span class="pre">--ports_num_for_sparse</span></code>参数</li>
<li>CONF_PADDLE_GRADIENT_NUM：训练节点数量，即<code class="docutils literal"><span class="pre">--num_gradient_servers参数</span></code></li>
</ul>
<p>这些参数的具体描述，读者可以查看<a class="reference external" href="http://www.paddlepaddle.org/docs/develop/documentation/zh/howto/usage/cmd_parameter/detail_introduction_cn.html">这里</a>。</p>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 
<p>编写完YAML文件后，可以使用Kubernetes的命令行工具创建job。</p>
<div class="highlight-bash"><div class="highlight"><pre><span></span>kubectl create -f job.yaml
</pre></div>
</div>
<p>创建成功后，Kubernetes就会创建3个pod作为PaddlePaddle节点然后拉取镜像，启动容器开始训练。</p>
</div>
<div class="section" id="">
<span id="id4"></span><h3>查看输出<a class="headerlink" href="#" title="永久链接至标题"></a></h3>
<p>在训练过程中，可以在共享存储上查看输出的日志和模型，例如output目录下就存放了输出结果。注意node_0，node_1，node_2这几个目录表示PaddlePaddle节点与trainer_id，并不是Kubernetes中的node概念。</p>
<div class="highlight-bash"><div class="highlight"><pre><span></span><span class="o">[</span>root@paddle-kubernetes-node0 output<span class="o">]</span><span class="c1"># tree -d</span>
.
├── node_0
│   ├── server.log
│   └── train.log
├── node_1
│   ├── server.log
│   └── train.log
├── node_2
......
├── pass-00002
│   ├── <span class="k">done</span>
│   ├── ___embedding_0__.w0
│   ├── ___embedding_1__.w0
......
</pre></div>
</div>
<p>我们可以通过日志查看容器训练的情况，例如：</p>
<div class="highlight-bash"><div class="highlight"><pre><span></span><span class="o">[</span>root@paddle-kubernetes-node0 node_0<span class="o">]</span><span class="c1"># cat train.log</span>
I1116 <span class="m">09</span>:10:17.123121    <span class="m">50</span> Util.cpp:155<span class="o">]</span> commandline:
 /usr/local/bin/../opt/paddle/bin/paddle_trainer
    --nics<span class="o">=</span>eth0 --port<span class="o">=</span><span class="m">7164</span>
    --ports_num<span class="o">=</span><span class="m">2</span> --comment<span class="o">=</span>paddle_process_by_paddle
    --pservers<span class="o">=</span><span class="m">192</span>.168.129.66,192.168.223.143,192.168.129.71
    --ports_num_for_sparse<span class="o">=</span><span class="m">2</span> --config<span class="o">=</span>./trainer_config.py
    --trainer_count<span class="o">=</span><span class="m">4</span> --num_passes<span class="o">=</span><span class="m">10</span> --use_gpu<span class="o">=</span><span class="m">0</span>
    --log_period<span class="o">=</span><span class="m">50</span> --dot_period<span class="o">=</span><span class="m">10</span> --saving_period<span class="o">=</span><span class="m">1</span>
    --local<span class="o">=</span><span class="m">0</span> --trainer_id<span class="o">=</span><span class="m">0</span>
    --save_dir<span class="o">=</span>/home/jobpath/paddle-cluster-job/output
I1116 <span class="m">09</span>:10:17.123440    <span class="m">50</span> Util.cpp:130<span class="o">]</span> Calling runInitFunctions
I1116 <span class="m">09</span>:10:17.123764    <span class="m">50</span> Util.cpp:143<span class="o">]</span> Call runInitFunctions <span class="k">done</span>.
<span class="o">[</span>WARNING <span class="m">2016</span>-11-16 <span class="m">09</span>:10:17,227 default_decorators.py:40<span class="o">]</span> please use keyword arguments in paddle config.
<span class="o">[</span>INFO <span class="m">2016</span>-11-16 <span class="m">09</span>:10:17,239 networks.py:1282<span class="o">]</span> The input order is <span class="o">[</span>movie_id, title, genres, user_id, gender, age, occupation, rating<span class="o">]</span>
T
Deploy to GitHub Pages: d7b2058433589619e041e20016853c7ee2ec10b6  
Travis CI 已提交
417 
<span class="o">[</span>INFO <span class="m">2016</span>-11-16 <span class="m">09</span>:10:17,239 networks.py:1289<span class="o">]</span> The output order is <span class="o">[</span>__square_error_cost_0__<span class="o">]</span>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 
I1116 <span class="m">09</span>:10:17.392917    <span class="m">50</span> Trainer.cpp:170<span class="o">]</span> trainer mode: Normal
I1116 <span class="m">09</span>:10:17.613910    <span class="m">50</span> PyDataProvider2.cpp:257<span class="o">]</span> loading dataprovider dataprovider::process
I1116 <span class="m">09</span>:10:17.680917    <span class="m">50</span> PyDataProvider2.cpp:257<span class="o">]</span> loading dataprovider dataprovider::process
I1116 <span class="m">09</span>:10:17.681543    <span class="m">50</span> GradientMachine.cpp:134<span class="o">]</span> Initing parameters..
I1116 <span class="m">09</span>:10:18.012390    <span class="m">50</span> GradientMachine.cpp:141<span class="o">]</span> Init parameters <span class="k">done</span>.
I1116 <span class="m">09</span>:10:18.018641    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">0</span> <span class="m">192</span>.168.129.66:7164
I1116 <span class="m">09</span>:10:18.018950    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">1</span> <span class="m">192</span>.168.129.66:7165
I1116 <span class="m">09</span>:10:18.019069    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">2</span> <span class="m">192</span>.168.223.143:7164
I1116 <span class="m">09</span>:10:18.019492    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">3</span> <span class="m">192</span>.168.223.143:7165
I1116 <span class="m">09</span>:10:18.019716    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">4</span> <span class="m">192</span>.168.129.71:7164
I1116 <span class="m">09</span>:10:18.019836    <span class="m">50</span> ParameterClient2.cpp:122<span class="o">]</span> pserver <span class="m">5</span> <span class="m">192</span>.168.129.71:7165
</pre></div>
</div>
</div>
</div>
<div class="section" id="">
<span id="id5"></span><h2>一些细节的补充<a class="headerlink" href="#" title="永久链接至标题"></a></h2>
<div class="section" id="">
<span id="id6"></span><h3>使用环境变量<a class="headerlink" href="#" title="永久链接至标题"></a></h3>
<p>使用容器方式运行训练任务的Kubernetes Job，通常会使用环境变量配置Job的配置信息<code class="docutils literal"><span class="pre">start_paddle.py</span></code>提供了一个启动脚本，将环境变量转换成paddle的命令行参数：</p>
<div class="highlight-default"><div class="highlight"><pre><span></span><span class="n">API</span> <span class="o">=</span> <span class="s2">&quot;/api/v1/namespaces/&quot;</span>
<span class="n">JOBSELECTOR</span> <span class="o">=</span> <span class="s2">&quot;labelSelector=job-name=&quot;</span>
<span class="n">JOB_PATH</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;JOB_PATH&quot;</span><span class="p">)</span> <span class="o">+</span> <span class="s2">&quot;/&quot;</span> <span class="o">+</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;JOB_NAME&quot;</span><span class="p">)</span>
<span class="n">JOB_PATH_OUTPUT</span> <span class="o">=</span> <span class="n">JOB_PATH</span> <span class="o">+</span> <span class="s2">&quot;/output&quot;</span>
<span class="n">JOBNAME</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;JOB_NAME&quot;</span><span class="p">)</span>
<span class="n">NAMESPACE</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;JOB_NAMESPACE&quot;</span><span class="p">)</span>
<span class="n">PADDLE_NIC</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;CONF_PADDLE_NIC&quot;</span><span class="p">)</span>
<span class="n">PADDLE_PORT</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;CONF_PADDLE_PORT&quot;</span><span class="p">)</span>
<span class="n">PADDLE_PORTS_NUM</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;CONF_PADDLE_PORTS_NUM&quot;</span><span class="p">)</span>
<span class="n">PADDLE_PORTS_NUM_SPARSE</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;CONF_PADDLE_PORTS_NUM_SPARSE&quot;</span><span class="p">)</span>
<span class="n">PADDLE_SERVER_NUM</span> <span class="o">=</span> <span class="n">os</span><span class="o">.</span><span class="n">getenv</span><span class="p">(</span><span class="s2">&quot;CONF_PADDLE_GRADIENT_NUM&quot;</span><span class="p">)</span>
</pre></div>
</div>
</div>
<div class="section" id="pod">
<span id="pod"></span><h3>Pod间通信<a class="headerlink" href="#pod" title="永久链接至标题"></a></h3>
<p><code class="docutils literal"><span class="pre">start_paddle.py</span></code>脚本开始时，会先进行参数的初始化与解析。</p>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="n">parser</span> <span class="o">=</span> <span class="n">argparse</span><span class="o">.</span><span class="n">ArgumentParser</span><span class="p">(</span><span class="n">prog</span><span class="o">=</span><span class="s2">&quot;start_paddle.py&quot;</span><span class="p">,</span>
                                     <span class="n">description</span><span class="o">=</span><span class="s1">&#39;simple tool for k8s&#39;</span><span class="p">)</span>
    <span class="n">args</span><span class="p">,</span> <span class="n">train_args_list</span> <span class="o">=</span> <span class="n">parser</span><span class="o">.</span><span class="n">parse_known_args</span><span class="p">()</span>
    <span class="n">train_args</span> <span class="o">=</span> <span class="n">refine_unknown_args</span><span class="p">(</span><span class="n">train_args_list</span><span class="p">)</span>
    <span class="n">train_args_dict</span> <span class="o">=</span> <span class="nb">dict</span><span class="p">(</span><span class="nb">zip</span><span class="p">(</span><span class="n">train_args</span><span class="p">[:</span><span class="o">-</span><span class="mi">1</span><span class="p">:</span><span class="mi">2</span><span class="p">],</span> <span class="n">train_args</span><span class="p">[</span><span class="mi">1</span><span class="p">::</span><span class="mi">2</span><span class="p">]))</span>
    <span class="n">podlist</span> <span class="o">=</span> <span class="n">getPodList</span><span class="p">()</span>
</pre></div>
</div>
<p>然后通过函数<code class="docutils literal"><span class="pre">getPodList()</span></code>访问Kubernetes的接口来查询此job对应的所有pod信息。当所有pod都处于running状态（容器运行都运行）时，再通过函数<code class="docutils literal"><span class="pre">getIdMap(podlist)</span></code>获取trainer_id。</p>
<div class="highlight-python"><div class="highlight"><pre><span></span>    <span class="n">podlist</span> <span class="o">=</span> <span class="n">getPodList</span><span class="p">()</span>
    <span class="c1"># need to wait until all pods are running</span>
    <span class="k">while</span> <span class="ow">not</span> <span class="n">isPodAllRunning</span><span class="p">(</span><span class="n">podlist</span><span class="p">):</span>
        <span class="n">time</span><span class="o">.</span><span class="n">sleep</span><span class="p">(</span><span class="mi">10</span><span class="p">)</span>
        <span class="n">podlist</span> <span class="o">=</span> <span class="n">getPodList</span><span class="p">()</span>
    <span class="n">idMap</span> <span class="o">=</span> <span class="n">getIdMap</span><span class="p">(</span><span class="n">podlist</span><span class="p">)</span>
</pre></div>
</div>
<ul class="simple">
<li><em>注意</em>: <code class="docutils literal"><span class="pre">getPodList()</span></code>会获取当前namespace下的所有pod，如果已经有pod运行，可能会导致出错。这种集群节点管理方式会在将来使用<a class="reference external" href="https://kubernetes.io/docs/concepts/abstractions/controllers/statefulsets/">statfulsets</a>代替。</li>
</ul>
<p>在函数<code class="docutils literal"><span class="pre">getIdMap(podlist)</span></code>内部，我们通过读取<code class="docutils literal"><span class="pre">podlist</span></code>中每个pod的IP地址，将IP排序生成的序号作为trainer_id。</p>
<div class="highlight-python"><div class="highlight"><pre><span></span><span class="k">def</span> <span class="nf">getIdMap</span><span class="p">(</span><span class="n">podlist</span><span class="p">):</span>
    <span class="sd">&#39;&#39;&#39;</span>
<span class="sd">    generate tainer_id by ip</span>
<span class="sd">    &#39;&#39;&#39;</span>
    <span class="n">ips</span> <span class="o">=</span> <span class="p">[]</span>
    <span class="k">for</span> <span class="n">pod</span> <span class="ow">in</span> <span class="n">podlist</span><span class="p">[</span><span class="s2">&quot;items&quot;</span><span class="p">]:</span>
        <span class="n">ips</span><span class="o">.</span><span class="n">append</span><span class="p">(</span><span class="n">pod</span><span class="p">[</span><span class="s2">&quot;status&quot;</span><span class="p">][</span><span class="s2">&quot;podIP&quot;</span><span class="p">])</span>
    <span class="n">ips</span><span class="o">.</span><span class="n">sort</span><span class="p">()</span>
    <span class="n">idMap</span> <span class="o">=</span> <span class="p">{}</span>
    <span class="k">for</span> <span class="n">i</span> <span class="ow">in</span> <span class="nb">range</span><span class="p">(</span><span class="nb">len</span><span class="p">(</span><span class="n">ips</span><span class="p">)):</span>
        <span class="n">idMap</span><span class="p">[</span><span class="n">ips</span><span class="p">[</span><span class="n">i</span><span class="p">]]</span> <span class="o">=</span> <span class="n">i</span>
    <span class="k">return</span> <span class="n">idMap</span>
</pre></div>
</div>
<p>在得到<code class="docutils literal"><span class="pre">idMap</span></code>后，通过函数<code class="docutils literal"><span class="pre">startPaddle(idMap,</span> <span class="pre">train_args_dict)</span></code>构造<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>与<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>的启动参数并执行进程。</p>
</div>
<div class="section" id="">
<span id="id7"></span><h3>启动任务<a class="headerlink" href="#" title="永久链接至标题"></a></h3>
<p>在函数<code class="docutils literal"><span class="pre">startPaddle</span></code>中，最主要的工作就是解析出<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">pserver</span></code>与<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>的启动参数。例如<code class="docutils literal"><span class="pre">paddle</span> <span class="pre">train</span></code>参数的解析，解析环境变量得到<code class="docutils literal"><span class="pre">PADDLE_NIC</span></code>，<code class="docutils literal"><span class="pre">PADDLE_PORT</span></code>，<code class="docutils literal"><span class="pre">PADDLE_PORTS_NUM</span></code>等参数，然后通过自身的IP地址在<code class="docutils literal"><span class="pre">idMap</span></code>中获取<code class="docutils literal"><span class="pre">trainerId</span></code>。</p>
<div class="highlight-python"><div class="highlight"><pre><span></span>    <span class="n">program</span> <span class="o">=</span> <span class="s1">&#39;paddle train&#39;</span>
    <span class="n">args</span> <span class="o">=</span> <span class="s2">&quot; --nics=&quot;</span> <span class="o">+</span> <span class="n">PADDLE_NIC</span>
    <span class="n">args</span> <span class="o">+=</span> <span class="s2">&quot; --port=&quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">PADDLE_PORT</span><span class="p">)</span>
    <span class="n">args</span> <span class="o">+=</span> <span class="s2">&quot; --ports_num=&quot;</span> <span class="o">+</span> <span class="nb">str</span><span class="p">(</span><span class="n">PADDLE_PORTS_NUM</span><span class="p">)</span>
    <span class="n">args</span> <span class="o">+=</span> <span class="s2">&quot; --comment=&quot;</span> <span class="o">+</span> <span class="s2">&quot;paddle_process_by_paddle&quot;</span>
    <span class="n">ip_string</span> <span class="o">=</span> <span class="s2">&quot;&quot;</span>
    <span class="k">for</span> <span class="n">ip</span> <span class="ow">in</span> <span class="n">idMap</span><span class="o">.</span><span class="n">keys</span><span class="p">():</span>
        <span class="n">ip_string</span> <span class="o">+=</span> <span class="p">(</span><span class="n">ip</span> <span class="o">+</span> <span class="s2">&quot;,&quot;</span><span class="p">)</span>
    <span class="n">ip_string</span> <span class="o">=</span> <span class="n">ip_string</span><span class="o">.</span><span class="n">rstrip</span><span class="p">(</span><span class="s2">&quot;,&quot;</span><span class="p">)</span>
    <span class="n">args</span> <span class="o">+=</span> <span class="s2">&quot; --pservers=&quot;</span> <span class="o">+</span> <span class="n">ip_string</span>
    <span class="n">args_ext</span> <span class="o">=</span> <span class="s2">&quot;&quot;</span>
    <span class="k">for</span> <span class="n">key</span><span class="p">,</span> <span class="n">value</span> <span class="ow">in</span> <span class="n">train_args_dict</span><span class="o">.</span><span class="n">items</span><span class="p">():</span>
        <span class="n">args_ext</span> <span class="o">+=</span> <span class="p">(</span><span class="s1">&#39; --&#39;</span> <span class="o">+</span> <span class="n">key</span> <span class="o">+</span> <span class="s1">&#39;=&#39;</span> <span class="o">+</span> <span class="n">value</span><span class="p">)</span>
    <span class="n">localIP</span> <span class="o">=</span> <span class="n">socket</span><span class="o">.</span><span class="n">gethostbyname</span><span class="p">(</span><span class="n">socket</span><span class="o">.</span><span class="n">gethostname</span><span class="p">())</span>
    <span class="n">trainerId</span> <span class="o">=</span> <span class="n">idMap</span><span class="p">[</span><span class="n">localIP</span><span class="p">]</span>
    <span class="n">args</span> <span class="o">+=</span> <span class="s2">&quot; &quot;</span> <span class="o">+</span> <span class="n">args_ext</span> <span class="o">+</span> <span class="s2">&quot; --trainer_id=&quot;</span> <span class="o">+</span> \
        <span class="nb">str</span><span class="p">(</span><span class="n">trainerId</span><span class="p">)</span> <span class="o">+</span> <span class="s2">&quot; --save_dir=&quot;</span> <span class="o">+</span> <span class="n">JOB_PATH_OUTPUT</span>
</pre></div>
</div>
</div>
</div>
</div>


           </div>
          </div>
          <footer>
  
    <div class="rst-footer-buttons" role="navigation" aria-label="footer navigation">
T
Deploy to GitHub Pages: 16fd9f18e304043a1035ad237884a93be354804c  
Travis CI 已提交
525 
        <a href="k8s_aws_cn.html" class="btn btn-neutral float-right" title="Distributed PaddlePaddle Training on AWS with Kubernetes" accesskey="n">Next <span class="fa fa-arrow-circle-right"></span></a>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 
      
      
        <a href="k8s_cn.html" class="btn btn-neutral" title="Kubernetes单机训练" accesskey="p"><span class="fa fa-arrow-circle-left"></span> Previous</a>
      
    </div>
  

  <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',
T
Deploy to GitHub Pages: 18e5edf6a5b09868dbb083c312fe852cb8c29e55  
Travis CI 已提交
562 563 
            HAS_SOURCE:  true,
            SOURCELINK_SUFFIX: ".txt",
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
564 565 566 567 568 569 
        };
    </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="../../../_static/translations.js"></script>
T
Deploy to GitHub Pages: 93007fea7f57cbb3d1297a7628d5604374393096  
Travis CI 已提交
570 
      <script type="text/javascript" src="https://cdn.bootcss.com/mathjax/2.7.0/MathJax.js"></script>
T
Deploy to GitHub Pages: 2c54905c3e3154327a8eb57ad96d0770dcc5f435  
Travis CI 已提交
571 572 573 574 575 576 577 578 579 580 581 582 583 
       
  

  
  
    <script type="text/javascript" src="../../../_static/js/theme.js"></script>
  
  
  <script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.7/js/bootstrap.min.js" integrity="sha384-Tc5IQib027qvyjSMfHjOMaLkfuWVxZxUPnCJA7l2mCWNIpG9mGCD8wGNIcPD7Txa" crossorigin="anonymous"></script>
  <script src="https://cdn.jsdelivr.net/perfect-scrollbar/0.6.14/js/perfect-scrollbar.jquery.min.js"></script>
  <script src="../../../_static/js/paddle_doc_init.js"></script> 

</body>
T
Deploy to GitHub Pages: 299e959a190017689d18af70d300f681745f4c67  
Travis CI 已提交
584 
</html>