提交 7a346923 编写于 作者: T Travis CI

Deploy to GitHub Pages: c52a0bde

上级 db646c0d
# Design Doc: NCCL support in Paddle Fluid
## Abstract
This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script.
## Motivation
[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel.
- Pros
1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
1. high performance in NVIDIA GPUs.
1. MPI like primitives, which have low learning cost for users.
- Cons
1. Only design for NVIDIA GPUs, not a general multi-device solution.
1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.
At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.
As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.
Besides, it needs interfaces to synchronize model update with each different GPU Cards.
## Implementation
As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
### Transpiler
To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.
1. The user-defined model will be a single device program
2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator.
*Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
<img src="images/multigpu_before_convert.png" width="300"/>
After compiling, the graph as shows
<img src="images/multigpu_allreduce.png" width="1000"/>
Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc.
- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU.
Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.
As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
- **AllReduce**
Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is
1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.
2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.
4. Then the root card will optimize the parameter.
5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.
6. Finish the sychronization round.
The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.
<!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: NCCL support in Paddle Fluid &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"/>
<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">
<a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Fork me on Github</a>
<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">
<li><a href="/">Home</a></li>
</ul>
</div>
<div class="doc-module">
<ul>
<li class="toctree-l1"><a class="reference internal" href="../getstarted/index_en.html">GET STARTED</a></li>
<li class="toctree-l1"><a class="reference internal" href="../howto/index_en.html">HOW TO</a></li>
<li class="toctree-l1"><a class="reference internal" href="../api/index_en.html">API</a></li>
<li class="toctree-l1"><a class="reference internal" href="../mobile/index_en.html">MOBILE</a></li>
</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>
<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/build_and_install/index_en.html">Install and Build</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../getstarted/build_and_install/pip_install_en.html">Install Using pip</a></li>
<li class="toctree-l3"><a class="reference internal" href="../getstarted/build_and_install/docker_install_en.html">Run in Docker Containers</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/dev/build_en.html">Build using Docker</a></li>
<li class="toctree-l3"><a class="reference internal" href="../getstarted/build_and_install/build_from_source_en.html">Build from Sources</a></li>
</ul>
</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/usage/cmd_parameter/index_en.html">Set Command-line Parameters</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/use_case_en.html">Use Case</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/arguments_en.html">Argument Outline</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/detail_introduction_en.html">Detail Description</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/cluster/cluster_train_en.html">PaddlePaddle Distributed Training</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/k8s/k8s_en.html">Paddle On Kubernetes</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/k8s/k8s_aws_en.html">Distributed PaddlePaddle Training on AWS with Kubernetes</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/dev/new_layer_en.html">Write New Layers</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/dev/contribute_to_paddle_en.html">Contribute Code</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/dev/write_docs_en.html">Contribute Documentation</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/deep_model/rnn/index_en.html">RNN Models</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/deep_model/rnn/rnn_config_en.html">RNN Configuration</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="../api/index_en.html">API</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../api/v2/model_configs.html">Model Configuration</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>
<li class="toctree-l3"><a class="reference internal" href="../api/v2/config/evaluators.html">Evaluators</a></li>
<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>
<li class="toctree-l2"><a class="reference internal" href="../api/v2/data.html">Data Reader Interface and DataSets</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>
<li class="toctree-l2"><a class="reference internal" href="../api/v2/run_logic.html">Training and Inference</a></li>
<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>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../mobile/index_en.html">MOBILE</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../mobile/cross_compiling_for_android_en.html">Build PaddlePaddle for Android</a></li>
<li class="toctree-l2"><a class="reference internal" href="../mobile/cross_compiling_for_raspberry_en.html">Build PaddlePaddle for Raspberry Pi</a></li>
</ul>
</li>
</ul>
</nav>
<section class="doc-content-wrap">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="wy-breadcrumbs">
<li>Design Doc: NCCL support in Paddle Fluid</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="design-doc-nccl-support-in-paddle-fluid">
<span id="design-doc-nccl-support-in-paddle-fluid"></span><h1>Design Doc: NCCL support in Paddle Fluid<a class="headerlink" href="#design-doc-nccl-support-in-paddle-fluid" title="Permalink to this headline"></a></h1>
<div class="section" id="abstract">
<span id="abstract"></span><h2>Abstract<a class="headerlink" href="#abstract" title="Permalink to this headline"></a></h2>
<p>This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives <code class="docutils literal"><span class="pre">Broadcast</span></code>, <code class="docutils literal"><span class="pre">Allreduce</span></code>, <code class="docutils literal"><span class="pre">Reduce</span></code> as operators to utilize Multi-GPU powers in one script.</p>
</div>
<div class="section" id="motivation">
<span id="motivation"></span><h2>Motivation<a class="headerlink" href="#motivation" title="Permalink to this headline"></a></h2>
<p><a class="reference external" href="https://developer.nvidia.com/nccl">NCCL</a> is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel.</p>
<ul class="simple">
<li>Pros</li>
</ul>
<ol class="simple">
<li>easily plug-in with <a class="reference external" href="https://developer.nvidia.com/nccl">NCCL2</a> library.</li>
<li>high performance in NVIDIA GPUs.</li>
<li>MPI like primitives, which have low learning cost for users.</li>
</ol>
<ul class="simple">
<li>Cons</li>
</ul>
<ol class="simple">
<li>Only design for NVIDIA GPUs, not a general multi-device solution.</li>
<li>Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.</li>
</ol>
<p>At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.</p>
<p>As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.</p>
<p>Besides, it needs interfaces to synchronize model update with each different GPU Cards.</p>
</div>
<div class="section" id="implementation">
<span id="implementation"></span><h2>Implementation<a class="headerlink" href="#implementation" title="Permalink to this headline"></a></h2>
<p>As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.</p>
<div class="section" id="transpiler">
<span id="transpiler"></span><h3>Transpiler<a class="headerlink" href="#transpiler" title="Permalink to this headline"></a></h3>
<p>To be compatible with <a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md">parameter server design doc</a>, the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.</p>
<ol>
<li><p class="first">The user-defined model will be a single device program</p>
</li>
<li><p class="first">Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the <code class="docutils literal"><span class="pre">Send</span></code>, <code class="docutils literal"><span class="pre">Recv</span></code> operator.</p>
<p><em>Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, <a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter">Send, Recv</a> in multiple machines</em></p>
<p><img src="images/multigpu_before_convert.png" width="300"/></p>
</li>
</ol>
<p>After compiling, the graph as shows</p>
<p><img src="images/multigpu_allreduce.png" width="1000"/></p>
<p>Operators are added to the sub-graphs. Every GPU assigned a role of <code class="docutils literal"><span class="pre">rank0</span></code>, <code class="docutils literal"><span class="pre">rank1</span></code> etc.</p>
<ul class="simple">
<li><strong>Broadcast</strong>. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from<code class="docutils literal"><span class="pre">rank0</span></code> GPU.</li>
<li><strong>AllReduce</strong>. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU.</li>
</ul>
<p>Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.</p>
<p>As it shown in the picture, when each GPU compute the gradient of <code class="docutils literal"><span class="pre">W</span></code>, followed with a <code class="docutils literal"><span class="pre">AllReduce</span></code> operator, accumulate the <code class="docutils literal"><span class="pre">dW</span></code> to full batch of data, then run the optimize process individually and apply the gradient to its <code class="docutils literal"><span class="pre">W</span></code>.</p>
<ul class="simple">
<li><strong>AllReduce</strong>
Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is</li>
</ul>
<ol class="simple">
<li>Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.</li>
<li>The whole model&#8217;s parameter will be hashed to different root card, ensure the load balance between GPUs.</li>
<li>Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.</li>
<li>Then the root card will optimize the parameter.</li>
<li>This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.</li>
<li>Finish the sychronization round.</li>
</ol>
<p>The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.</p>
</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,
SOURCELINK_SUFFIX: ".txt",
};
</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 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>
</html>
\ No newline at end of file
因为 它太大了无法显示 source diff 。你可以改为 查看blob
# Design Doc: NCCL support in Paddle Fluid
## Abstract
This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives `Broadcast`, `Allreduce`, `Reduce` as operators to utilize Multi-GPU powers in one script.
## Motivation
[NCCL](https://developer.nvidia.com/nccl) is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel.
- Pros
1. easily plug-in with [NCCL2](https://developer.nvidia.com/nccl) library.
1. high performance in NVIDIA GPUs.
1. MPI like primitives, which have low learning cost for users.
- Cons
1. Only design for NVIDIA GPUs, not a general multi-device solution.
1. Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.
At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.
As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.
Besides, it needs interfaces to synchronize model update with each different GPU Cards.
## Implementation
As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.
### Transpiler
To be compatible with [parameter server design doc](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md), the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.
1. The user-defined model will be a single device program
2. Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the `Send`, `Recv` operator.
*Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, [Send, Recv](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter) in multiple machines*
<img src="images/multigpu_before_convert.png" width="300"/>
After compiling, the graph as shows
<img src="images/multigpu_allreduce.png" width="1000"/>
Operators are added to the sub-graphs. Every GPU assigned a role of `rank0`, `rank1` etc.
- **Broadcast**. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from`rank0` GPU.
- **AllReduce**. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU.
Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.
As it shown in the picture, when each GPU compute the gradient of `W`, followed with a `AllReduce` operator, accumulate the `dW` to full batch of data, then run the optimize process individually and apply the gradient to its `W`.
- **AllReduce**
Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is
1. Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.
2. The whole model's parameter will be hashed to different root card, ensure the load balance between GPUs.
3. Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.
4. Then the root card will optimize the parameter.
5. This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.
6. Finish the sychronization round.
The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.
<!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: NCCL support in Paddle Fluid &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"/>
<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">
<a class="fork-on-github" href="https://github.com/PaddlePaddle/Paddle" target="_blank"><i class="fa fa-github"></i>Fork me on Github</a>
<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">
<li><a href="/">Home</a></li>
</ul>
</div>
<div class="doc-module">
<ul>
<li class="toctree-l1"><a class="reference internal" href="../getstarted/index_cn.html">新手入门</a></li>
<li class="toctree-l1"><a class="reference internal" href="../howto/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>
<li class="toctree-l1"><a class="reference internal" href="../mobile/index_cn.html">MOBILE</a></li>
</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>
<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>
<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>
<li class="toctree-l3"><a class="reference internal" href="../howto/dev/build_cn.html">用Docker编译和测试PaddlePaddle</a></li>
<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>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../getstarted/concepts/use_concepts_cn.html">基本使用概念</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../howto/index_cn.html">进阶指南</a><ul>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/cmd_parameter/index_cn.html">设置命令行参数</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/use_case_cn.html">使用案例</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/arguments_cn.html">参数概述</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/usage/cmd_parameter/detail_introduction_cn.html">细节描述</a></li>
</ul>
</li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/cluster/cluster_train_cn.html">PaddlePaddle分布式训练</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/k8s/k8s_basis_cn.html">Kubernetes 简介</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/k8s/k8s_cn.html">Kubernetes单机训练</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/usage/k8s/k8s_distributed_cn.html">Kubernetes分布式训练</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/dev/contribute_to_paddle_cn.html">如何贡献代码</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/dev/write_docs_cn.html">如何贡献/修改文档</a></li>
<li class="toctree-l2"><a class="reference internal" href="../howto/deep_model/rnn/index_cn.html">RNN相关模型</a><ul>
<li class="toctree-l3"><a class="reference internal" href="../howto/deep_model/rnn/rnn_config_cn.html">RNN配置</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/deep_model/rnn/recurrent_group_cn.html">Recurrent Group教程</a></li>
<li class="toctree-l3"><a class="reference internal" href="../howto/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="../howto/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="../howto/optimization/gpu_profiling_cn.html">GPU性能分析与调优</a></li>
</ul>
</li>
<li class="toctree-l1"><a class="reference internal" href="../api/index_cn.html">API</a><ul>
<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>
<li class="toctree-l3"><a class="reference internal" href="../api/v2/config/evaluators.html">Evaluators</a></li>
<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>
<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>
<li class="toctree-l2"><a class="reference internal" href="../api/v2/run_logic.html">训练与应用</a></li>
</ul>
</li>
<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>
<li class="toctree-l1"><a class="reference internal" href="../mobile/index_cn.html">MOBILE</a><ul>
<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>
</ul>
</li>
</ul>
</nav>
<section class="doc-content-wrap">
<div role="navigation" aria-label="breadcrumbs navigation">
<ul class="wy-breadcrumbs">
<li>Design Doc: NCCL support in Paddle Fluid</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="design-doc-nccl-support-in-paddle-fluid">
<span id="design-doc-nccl-support-in-paddle-fluid"></span><h1>Design Doc: NCCL support in Paddle Fluid<a class="headerlink" href="#design-doc-nccl-support-in-paddle-fluid" title="永久链接至标题"></a></h1>
<div class="section" id="abstract">
<span id="abstract"></span><h2>Abstract<a class="headerlink" href="#abstract" title="永久链接至标题"></a></h2>
<p>This Design Doc refers to the NCCL feature in paddle. We propose an approach to support NCCL library both on a single machine and multiple machines. We wrapper the NCCL primitives <code class="docutils literal"><span class="pre">Broadcast</span></code>, <code class="docutils literal"><span class="pre">Allreduce</span></code>, <code class="docutils literal"><span class="pre">Reduce</span></code> as operators to utilize Multi-GPU powers in one script.</p>
</div>
<div class="section" id="motivation">
<span id="motivation"></span><h2>Motivation<a class="headerlink" href="#motivation" title="永久链接至标题"></a></h2>
<p><a class="reference external" href="https://developer.nvidia.com/nccl">NCCL</a> is a NVIDIA library support Multi-GPU communicating and optimized for NVIDIA GPUs, it provides routines such as all-gather, all-reduce, broadcast, reduce, reduce-scatter, that can achieve high bandwidth over PCIe and NVLink high-speed interconnect. With NCCL library, we can easily accelerate the training in parallel.</p>
<ul class="simple">
<li>Pros</li>
</ul>
<ol class="simple">
<li>easily plug-in with <a class="reference external" href="https://developer.nvidia.com/nccl">NCCL2</a> library.</li>
<li>high performance in NVIDIA GPUs.</li>
<li>MPI like primitives, which have low learning cost for users.</li>
</ol>
<ul class="simple">
<li>Cons</li>
</ul>
<ol class="simple">
<li>Only design for NVIDIA GPUs, not a general multi-device solution.</li>
<li>Although NCCL1 is opensourced under BSD license, but NCCL2 is not opensourced anymore.</li>
</ol>
<p>At the beginning of training, the framework needs to distribute the same parameters to every GPU, and merge the gradients at any time user interests.</p>
<p>As a result, during training, we need the operations of peer to peer copy between different GPUs, aggregating gradients/parameters from GPUs, and broadcasting parameters to GPUs. Every GPU only need to run the operator with correct place information.</p>
<p>Besides, it needs interfaces to synchronize model update with each different GPU Cards.</p>
</div>
<div class="section" id="implementation">
<span id="implementation"></span><h2>Implementation<a class="headerlink" href="#implementation" title="永久链接至标题"></a></h2>
<p>As mentioned above, we wrap the NCCL routines as several kinds of operators. Need to note that NCCL need to create Communicator between gpu at the beginning, so there is a NCCLInit operator created.</p>
<div class="section" id="transpiler">
<span id="transpiler"></span><h3>Transpiler<a class="headerlink" href="#transpiler" title="永久链接至标题"></a></h3>
<p>To be compatible with <a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md">parameter server design doc</a>, the transpiler compiles the user defined operation graph into sub-graphs to be executed on different devices.</p>
<ol>
<li><p class="first">The user-defined model will be a single device program</p>
</li>
<li><p class="first">Broadcast/Reduce operators between GPUs will be inserted into the program, even for the multi-node, may insert the <code class="docutils literal"><span class="pre">Send</span></code>, <code class="docutils literal"><span class="pre">Recv</span></code> operator.</p>
<p><em>Broadcast, AllReduce in a single machine. And Broadcast, AllReduce, <a class="reference external" href="https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/ops/dist_train.md#graph-converter">Send, Recv</a> in multiple machines</em></p>
<p><img src="images/multigpu_before_convert.png" width="300"/></p>
</li>
</ol>
<p>After compiling, the graph as shows</p>
<p><img src="images/multigpu_allreduce.png" width="1000"/></p>
<p>Operators are added to the sub-graphs. Every GPU assigned a role of <code class="docutils literal"><span class="pre">rank0</span></code>, <code class="docutils literal"><span class="pre">rank1</span></code> etc.</p>
<ul class="simple">
<li><strong>Broadcast</strong>. Broadcast operator distribute initialized parameter to all the GPUs from the GPU who owns it. e.g. from<code class="docutils literal"><span class="pre">rank0</span></code> GPU.</li>
<li><strong>AllReduce</strong>. AllReduce operator synchronizes parameters/gradients between GPUs. AllReduce implemented in the Ring-Based communicating method, avoid of the bottle neck in a single GPU.</li>
</ul>
<p>Need to notice that AllReduce operator force GPUs synchronized at that point. The whole training process in asynchronous or synchronous mode depends on the AllReduce point in the graph.</p>
<p>As it shown in the picture, when each GPU compute the gradient of <code class="docutils literal"><span class="pre">W</span></code>, followed with a <code class="docutils literal"><span class="pre">AllReduce</span></code> operator, accumulate the <code class="docutils literal"><span class="pre">dW</span></code> to full batch of data, then run the optimize process individually and apply the gradient to its <code class="docutils literal"><span class="pre">W</span></code>.</p>
<ul class="simple">
<li><strong>AllReduce</strong>
Need to note that our AllReduce operator is a ring-base AllReduce implementation. If we use the NCCL2 AllReduce primitive, every GPU optimized full batch of data, wasted (n-1) GPU compute resources. In addition, NCCL2 built-in AllReduce will only utilize the communicating resource during synchronization, then update the gradient will be a subsequent phase. In fact, we can amortize the update gradient time cost into the communicating phase. The process is</li>
</ul>
<ol class="simple">
<li>Every parameter has its root card. That card will responsible for aggregating the gradients from GPUs.</li>
<li>The whole model&#8217;s parameter will be hashed to different root card, ensure the load balance between GPUs.</li>
<li>Logically neighberhood card will start send parameter to the next one. After one round, the parameter main card will aggregate the full gradients.</li>
<li>Then the root card will optimize the parameter.</li>
<li>This parameter card will send its optimized result to its neighberhood, then the neighberhood will send parameter to its next one.</li>
<li>Finish the sychronization round.</li>
</ol>
<p>The total time cost will be 2 * (n-1) * per-parameter-send-time, we reach the goal of amortize the upgrade time into communicating phase.</p>
</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,
SOURCELINK_SUFFIX: ".txt",
};
</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>
<script type="text/javascript" src="https://cdn.bootcss.com/mathjax/2.7.0/MathJax.js"></script>
<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>
</html>
\ No newline at end of file
因为 它太大了无法显示 source diff 。你可以改为 查看blob
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册