In [this slides](https://www.slideshare.net/cxwangyi/paddlepaddle-a-complete-solution-for-businesses), we explained that we'd like PaddlePaddle running on general-purpose clusters like those managed by Kubernetes, so to address demands for AI from both Internet and non-Internet industries.
This poses technical challenges to PaddlePaddle:
1. Support fault-recovery.
1. Support both offline and online training.
1. [Serverless computing](https://en.wikipedia.org/wiki/Serverless_computing) of distributed training.
## Training Job
A training job will be created once user asks Paddle cloud to train a model. The training job is made up of different processes that collaboratively consume data and produce a trained model. There are three kinds of processes:
1. the *master process*, which dispatches tasks to
1. one or more *trainer processes*, which run distributed training and synchronize gradients/models via
1. one or more *parameter server processes*, where each holds a shard of the global model.
Their relation is illustrated in the following graph:
<img src="src/paddle-model-sharding.png"/>
### Master Process
The master process will:
- Partition a dataset into [tasks](#task) and dispatch tasks to trainers.
- Keep track of training progress on the dataset with [task queue](#task-queue). A training job will iterate on the dataset for a full pass until it goes into next pass.
#### Task
A task is a data shard to be trained. The total number of tasks will be much bigger than the total number of trainers. The number of data instances inside a task will be much bigger than the mini-batch size.
#### Task Queue
The master process has three task queues to track training progress. As illustrated in the graph below, Job A and Job B both have one master process. Each master process has three task queues.
<img src="src/paddle-task-queues.png"/>
- The todo queue holds tasks to be dispatched. When a job starts, the master process fills in the todo queue with all tasks.
- The pending queue holds tasks that are currently training by trainers.
- the done queue holds tasks that are already trained.
The life cycle of a single task is illustrated below:
<img src="src/paddle-task-states.png"/>
1. When a new pass of training starts, all tasks will be placed in the todo queue.
1. The master process will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
1. The trainer will work on its tasks and tell the master process once a task is completed. The master process will dispatch a new task to that trainer.
1. If a task timeout. the master process will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
1. The master process will move completed task to the done queue. When the todo queue is empty, the master process will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
### Trainer Process
The trainer process will:
- Receive tasks from the master.
- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
### Parameter Server Process
Parameter server processes hold the parameters collaboratively. The parameters are partitioned on different parameter servers.
The parameter server will:
- Receive gradient from the trainers, update its parameters, and give the trainers the latest parameters.
- Periodically save its parameters to distributed file system by overriding the previous save.
### Optimization Algorithms
The communication pattern between the trainers and the parameter servers depends on the category of optimization algorithm:
Parameter server will wait for all trainer finish n-th mini-batch calculation and send their gradients before broadcasting new parameters to every trainer. Every trainer will wait for the new parameters before starting n+1-th mini-batch.
There will no synchronization between different trainers, and parameter server updates its parameter as soon as it receives new gradient:
- Each trainer uploads its accumulated gradient every n mini-batches.
- Every m mini-batches, the trainer downloads new parameters from parameter server.
- n and m do not have to be equal.
## Fault Tolerant
The training job will pause if the master processes is dead, or any of the parameter server process is dead. They will be started by [Kubernetes](https://kubernetes.io/) and recover in few minutes. Please refer to [fault recovery](#fault-recovery).
The training job will continue to make progress if there is at least one training process running. The strategy depends on the type of optimization algorithm:
- sync-SGD
TODO
- async-SGD
Since async-SGD does not require synchronization between mini-batches, the system will by definition make process if at least one trainer is running.
## Fault Recovery
PaddlePaddle uses [etcd](https://github.com/coreos/etcd) to keep track of the states of processes. Because etcd is a distributed reliable key-value store, the restarted process can recover its states from etcd. The model parameters are periodically saved into distributed file system, so a restarted parameter server can recover its parameters from the saved file.
Now we will introduce how each process recovers from a failure, the graph below shows how etcd is used:
<img src="src/paddle-etcd.png"/>
### Master Process
When the master is started by the Kubernetes, it executes the following steps at startup:
1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
The master process will kill itself if its etcd lease expires.
When the master process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
### Trainer Process
When the trainer is started by the Kubernetes, it executes the following steps at startup:
1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
1. Waits for tasks from the master to start training.
If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master process can discover the trainer again.
### Parameter Server Process
When the parameter server is started by Kubernetes, it executes the following steps at startup:
1. Read desired total number of parameter servers from etcd `/ps_desired`
1. Search through etcd keys `/ps/<index>` (`/ps/0`, `/ps/1`, ...) to find the first non-existant key whose index is smaller than the total number of parameter servers. Set the key using a transaction to avoid concurrent writes. The parameter server's index is inferred from the key name.
The desired number of parameter servers is 3:
<img src="src/paddle-ps-0.png"/>
The third parameter server joined:
<img src="src/paddle-ps-1.png"/>
1. The parameter server can load parameters if there are already saved parameters in the save path (inferred from its index).
1. Now the parameter server is ready for the trainers' requests.
If the parameter server's etcd lease expires, the parameter server will kill itself.
<liclass="toctree-l2"><aclass="reference internal"href="../../getstarted/build_and_install/index_en.html">Install and Build</a><ul>
<liclass="toctree-l3"><aclass="reference internal"href="../../getstarted/build_and_install/docker_install_en.html">PaddlePaddle in Docker Containers</a></li>
<liclass="toctree-l2"><aclass="reference internal"href="../../howto/usage/k8s/k8s_en.html">Paddle On Kubernetes</a></li>
<liclass="toctree-l2"><aclass="reference internal"href="../../howto/usage/k8s/k8s_aws_en.html">Distributed PaddlePaddle Training on AWS with Kubernetes</a></li>
<liclass="toctree-l2"><aclass="reference internal"href="../../howto/dev/new_layer_en.html">Write New Layers</a></li>
<spanid="design-doc-distributed-training"></span><h1>Design Doc: Distributed Training<aclass="headerlink"href="#design-doc-distributed-training"title="Permalink to this headline">¶</a></h1>
<divclass="section"id="objective">
<spanid="objective"></span><h2>Objective<aclass="headerlink"href="#objective"title="Permalink to this headline">¶</a></h2>
<p>In <aclass="reference external"href="https://www.slideshare.net/cxwangyi/paddlepaddle-a-complete-solution-for-businesses">this slides</a>, we explained that we’d like PaddlePaddle running on general-purpose clusters like those managed by Kubernetes, so to address demands for AI from both Internet and non-Internet industries.</p>
<p>This poses technical challenges to PaddlePaddle:</p>
<olclass="simple">
<li>Support fault-recovery.</li>
<li>Support both offline and online training.</li>
<li><aclass="reference external"href="https://en.wikipedia.org/wiki/Serverless_computing">Serverless computing</a> of distributed training.</li>
</ol>
</div>
<divclass="section"id="training-job">
<spanid="training-job"></span><h2>Training Job<aclass="headerlink"href="#training-job"title="Permalink to this headline">¶</a></h2>
<p>A training job will be created once user asks Paddle cloud to train a model. The training job is made up of different processes that collaboratively consume data and produce a trained model. There are three kinds of processes:</p>
<olclass="simple">
<li>the <em>master process</em>, which dispatches tasks to</li>
<li>one or more <em>trainer processes</em>, which run distributed training and synchronize gradients/models via</li>
<li>one or more <em>parameter server processes</em>, where each holds a shard of the global model.</li>
</ol>
<p>Their relation is illustrated in the following graph:</p>
<p><imgsrc="src/paddle-model-sharding.png"/></p>
<divclass="section"id="master-process">
<spanid="master-process"></span><h3>Master Process<aclass="headerlink"href="#master-process"title="Permalink to this headline">¶</a></h3>
<p>The master process will:</p>
<ulclass="simple">
<li>Partition a dataset into <aclass="reference external"href="#task">tasks</a> and dispatch tasks to trainers.</li>
<li>Keep track of training progress on the dataset with <aclass="reference external"href="#task-queue">task queue</a>. A training job will iterate on the dataset for a full pass until it goes into next pass.</li>
</ul>
<divclass="section"id="task">
<spanid="task"></span><h4>Task<aclass="headerlink"href="#task"title="Permalink to this headline">¶</a></h4>
<p>A task is a data shard to be trained. The total number of tasks will be much bigger than the total number of trainers. The number of data instances inside a task will be much bigger than the mini-batch size.</p>
</div>
<divclass="section"id="task-queue">
<spanid="task-queue"></span><h4>Task Queue<aclass="headerlink"href="#task-queue"title="Permalink to this headline">¶</a></h4>
<p>The master process has three task queues to track training progress. As illustrated in the graph below, Job A and Job B both have one master process. Each master process has three task queues.</p>
<p><imgsrc="src/paddle-task-queues.png"/></p>
<ulclass="simple">
<li>The todo queue holds tasks to be dispatched. When a job starts, the master process fills in the todo queue with all tasks.</li>
<li>The pending queue holds tasks that are currently training by trainers.</li>
<li>the done queue holds tasks that are already trained.</li>
</ul>
<p>The life cycle of a single task is illustrated below:</p>
<p><imgsrc="src/paddle-task-states.png"/></p>
<olclass="simple">
<li>When a new pass of training starts, all tasks will be placed in the todo queue.</li>
<li>The master process will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.</li>
<li>The trainer will work on its tasks and tell the master process once a task is completed. The master process will dispatch a new task to that trainer.</li>
<li>If a task timeout. the master process will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.</li>
<li>The master process will move completed task to the done queue. When the todo queue is empty, the master process will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.</li>
</ol>
</div>
</div>
<divclass="section"id="trainer-process">
<spanid="trainer-process"></span><h3>Trainer Process<aclass="headerlink"href="#trainer-process"title="Permalink to this headline">¶</a></h3>
<p>The trainer process will:</p>
<ulclass="simple">
<li>Receive tasks from the master.</li>
<li>Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.</li>
</ul>
</div>
<divclass="section"id="parameter-server-process">
<spanid="parameter-server-process"></span><h3>Parameter Server Process<aclass="headerlink"href="#parameter-server-process"title="Permalink to this headline">¶</a></h3>
<p>Parameter server processes hold the parameters collaboratively. The parameters are partitioned on different parameter servers.</p>
<p>The parameter server will:</p>
<ulclass="simple">
<li>Receive gradient from the trainers, update its parameters, and give the trainers the latest parameters.</li>
<li>Periodically save its parameters to distributed file system by overriding the previous save.</li>
</ul>
</div>
<divclass="section"id="optimization-algorithms">
<spanid="optimization-algorithms"></span><h3>Optimization Algorithms<aclass="headerlink"href="#optimization-algorithms"title="Permalink to this headline">¶</a></h3>
<p>The communication pattern between the trainers and the parameter servers depends on the category of optimization algorithm:</p>
<p>Parameter server will wait for all trainer finish n-th mini-batch calculation and send their gradients before broadcasting new parameters to every trainer. Every trainer will wait for the new parameters before starting n+1-th mini-batch.</p>
<p>There will no synchronization between different trainers, and parameter server updates its parameter as soon as it receives new gradient:</p>
<ulclass="simple">
<li>Each trainer uploads its accumulated gradient every n mini-batches.</li>
<li>Every m mini-batches, the trainer downloads new parameters from parameter server.</li>
<li>n and m do not have to be equal.</li>
</ul>
</li>
</ul>
</div>
</div>
<divclass="section"id="fault-tolerant">
<spanid="fault-tolerant"></span><h2>Fault Tolerant<aclass="headerlink"href="#fault-tolerant"title="Permalink to this headline">¶</a></h2>
<p>The training job will pause if the master processes is dead, or any of the parameter server process is dead. They will be started by <aclass="reference external"href="https://kubernetes.io/">Kubernetes</a> and recover in few minutes. Please refer to <aclass="reference external"href="#fault-recovery">fault recovery</a>.</p>
<p>The training job will continue to make progress if there is at least one training process running. The strategy depends on the type of optimization algorithm:</p>
<ul>
<li><pclass="first">sync-SGD</p>
<p>TODO</p>
</li>
<li><pclass="first">async-SGD</p>
<p>Since async-SGD does not require synchronization between mini-batches, the system will by definition make process if at least one trainer is running.</p>
</li>
</ul>
</div>
<divclass="section"id="fault-recovery">
<spanid="fault-recovery"></span><h2>Fault Recovery<aclass="headerlink"href="#fault-recovery"title="Permalink to this headline">¶</a></h2>
<p>PaddlePaddle uses <aclass="reference external"href="https://github.com/coreos/etcd">etcd</a> to keep track of the states of processes. Because etcd is a distributed reliable key-value store, the restarted process can recover its states from etcd. The model parameters are periodically saved into distributed file system, so a restarted parameter server can recover its parameters from the saved file.</p>
<p>Now we will introduce how each process recovers from a failure, the graph below shows how etcd is used:</p>
<p><imgsrc="src/paddle-etcd.png"/></p>
<divclass="section"id="master-process">
<spanid="id1"></span><h3>Master Process<aclass="headerlink"href="#master-process"title="Permalink to this headline">¶</a></h3>
<p>When the master is started by the Kubernetes, it executes the following steps at startup:</p>
<olclass="simple">
<li>Grabs a unique <em>master</em> lock in etcd, which prevents concurrent master instantiations.</li>
<li>Recovers the task queues from etcd if they already exist, otherwise, the master will create them.</li>
<li>Watches the trainer prefix keys <codeclass="docutils literal"><spanclass="pre">/trainer/</span></code> on etcd to find the live trainers.</li>
<li>Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.</li>
</ol>
<p>The master process will kill itself if its etcd lease expires.</p>
<p>When the master process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.</p>
</div>
<divclass="section"id="trainer-process">
<spanid="id2"></span><h3>Trainer Process<aclass="headerlink"href="#trainer-process"title="Permalink to this headline">¶</a></h3>
<p>When the trainer is started by the Kubernetes, it executes the following steps at startup:</p>
<olclass="simple">
<li>Watches the available parameter server prefix keys <codeclass="docutils literal"><spanclass="pre">/ps/</span></code> on etcd and waits until the count of parameter servers reaches the desired count.</li>
<li>Generates a unique ID, and sets key <codeclass="docutils literal"><spanclass="pre">/trainer/<unique</span><spanclass="pre">ID></span></code> with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.</li>
<li>Waits for tasks from the master to start training.</li>
</ol>
<p>If trainer’s etcd lease expires, it will try set key <codeclass="docutils literal"><spanclass="pre">/trainer/<unique</span><spanclass="pre">ID></span></code> again so that the master process can discover the trainer again.</p>
</div>
<divclass="section"id="parameter-server-process">
<spanid="id3"></span><h3>Parameter Server Process<aclass="headerlink"href="#parameter-server-process"title="Permalink to this headline">¶</a></h3>
<p>When the parameter server is started by Kubernetes, it executes the following steps at startup:</p>
<ol>
<li><pclass="first">Read desired total number of parameter servers from etcd <codeclass="docutils literal"><spanclass="pre">/ps_desired</span></code></p>
</li>
<li><pclass="first">Search through etcd keys <codeclass="docutils literal"><spanclass="pre">/ps/<index></span></code> (<codeclass="docutils literal"><spanclass="pre">/ps/0</span></code>, <codeclass="docutils literal"><spanclass="pre">/ps/1</span></code>, ...) to find the first non-existant key whose index is smaller than the total number of parameter servers. Set the key using a transaction to avoid concurrent writes. The parameter server’s index is inferred from the key name.</p>
<p>The desired number of parameter servers is 3:</p>
<p><imgsrc="src/paddle-ps-0.png"/></p>
<p>The third parameter server joined:</p>
<p><imgsrc="src/paddle-ps-1.png"/></p>
</li>
<li><pclass="first">The parameter server can load parameters if there are already saved parameters in the save path (inferred from its index).</p>
</li>
<li><pclass="first">Now the parameter server is ready for the trainers’ requests.</p>
</li>
</ol>
<p>If the parameter server’s etcd lease expires, the parameter server will kill itself.</p>
</div>
</div>
<divclass="section"id="dynamic-scaling">
<spanid="dynamic-scaling"></span><h2>Dynamic Scaling<aclass="headerlink"href="#dynamic-scaling"title="Permalink to this headline">¶</a></h2>
<divclass="section"id="trainer-scaling">
<spanid="trainer-scaling"></span><h3>Trainer Scaling<aclass="headerlink"href="#trainer-scaling"title="Permalink to this headline">¶</a></h3>
<p>TODO</p>
</div>
<divclass="section"id="parameter-server-scaling">
<spanid="parameter-server-scaling"></span><h3>Parameter Server Scaling<aclass="headerlink"href="#parameter-server-scaling"title="Permalink to this headline">¶</a></h3>
<p>Not planned for v1.</p>
</div>
</div>
<divclass="section"id="training-dataset-format">
<spanid="training-dataset-format"></span><h2>Training Dataset Format<aclass="headerlink"href="#training-dataset-format"title="Permalink to this headline">¶</a></h2>
<p>TODO</p>
</div>
<divclass="section"id="user-interface">
<spanid="user-interface"></span><h2>User Interface<aclass="headerlink"href="#user-interface"title="Permalink to this headline">¶</a></h2>
Built with <ahref="http://sphinx-doc.org/">Sphinx</a> using a <ahref="https://github.com/snide/sphinx_rtd_theme">theme</a> provided by <ahref="https://readthedocs.org">Read the Docs</a>.
In [this slides](https://www.slideshare.net/cxwangyi/paddlepaddle-a-complete-solution-for-businesses), we explained that we'd like PaddlePaddle running on general-purpose clusters like those managed by Kubernetes, so to address demands for AI from both Internet and non-Internet industries.
This poses technical challenges to PaddlePaddle:
1. Support fault-recovery.
1. Support both offline and online training.
1. [Serverless computing](https://en.wikipedia.org/wiki/Serverless_computing) of distributed training.
## Training Job
A training job will be created once user asks Paddle cloud to train a model. The training job is made up of different processes that collaboratively consume data and produce a trained model. There are three kinds of processes:
1. the *master process*, which dispatches tasks to
1. one or more *trainer processes*, which run distributed training and synchronize gradients/models via
1. one or more *parameter server processes*, where each holds a shard of the global model.
Their relation is illustrated in the following graph:
<img src="src/paddle-model-sharding.png"/>
### Master Process
The master process will:
- Partition a dataset into [tasks](#task) and dispatch tasks to trainers.
- Keep track of training progress on the dataset with [task queue](#task-queue). A training job will iterate on the dataset for a full pass until it goes into next pass.
#### Task
A task is a data shard to be trained. The total number of tasks will be much bigger than the total number of trainers. The number of data instances inside a task will be much bigger than the mini-batch size.
#### Task Queue
The master process has three task queues to track training progress. As illustrated in the graph below, Job A and Job B both have one master process. Each master process has three task queues.
<img src="src/paddle-task-queues.png"/>
- The todo queue holds tasks to be dispatched. When a job starts, the master process fills in the todo queue with all tasks.
- The pending queue holds tasks that are currently training by trainers.
- the done queue holds tasks that are already trained.
The life cycle of a single task is illustrated below:
<img src="src/paddle-task-states.png"/>
1. When a new pass of training starts, all tasks will be placed in the todo queue.
1. The master process will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.
1. The trainer will work on its tasks and tell the master process once a task is completed. The master process will dispatch a new task to that trainer.
1. If a task timeout. the master process will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.
1. The master process will move completed task to the done queue. When the todo queue is empty, the master process will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.
### Trainer Process
The trainer process will:
- Receive tasks from the master.
- Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.
### Parameter Server Process
Parameter server processes hold the parameters collaboratively. The parameters are partitioned on different parameter servers.
The parameter server will:
- Receive gradient from the trainers, update its parameters, and give the trainers the latest parameters.
- Periodically save its parameters to distributed file system by overriding the previous save.
### Optimization Algorithms
The communication pattern between the trainers and the parameter servers depends on the category of optimization algorithm:
Parameter server will wait for all trainer finish n-th mini-batch calculation and send their gradients before broadcasting new parameters to every trainer. Every trainer will wait for the new parameters before starting n+1-th mini-batch.
There will no synchronization between different trainers, and parameter server updates its parameter as soon as it receives new gradient:
- Each trainer uploads its accumulated gradient every n mini-batches.
- Every m mini-batches, the trainer downloads new parameters from parameter server.
- n and m do not have to be equal.
## Fault Tolerant
The training job will pause if the master processes is dead, or any of the parameter server process is dead. They will be started by [Kubernetes](https://kubernetes.io/) and recover in few minutes. Please refer to [fault recovery](#fault-recovery).
The training job will continue to make progress if there is at least one training process running. The strategy depends on the type of optimization algorithm:
- sync-SGD
TODO
- async-SGD
Since async-SGD does not require synchronization between mini-batches, the system will by definition make process if at least one trainer is running.
## Fault Recovery
PaddlePaddle uses [etcd](https://github.com/coreos/etcd) to keep track of the states of processes. Because etcd is a distributed reliable key-value store, the restarted process can recover its states from etcd. The model parameters are periodically saved into distributed file system, so a restarted parameter server can recover its parameters from the saved file.
Now we will introduce how each process recovers from a failure, the graph below shows how etcd is used:
<img src="src/paddle-etcd.png"/>
### Master Process
When the master is started by the Kubernetes, it executes the following steps at startup:
1. Grabs a unique *master* lock in etcd, which prevents concurrent master instantiations.
1. Recovers the task queues from etcd if they already exist, otherwise, the master will create them.
1. Watches the trainer prefix keys `/trainer/` on etcd to find the live trainers.
1. Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.
The master process will kill itself if its etcd lease expires.
When the master process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.
### Trainer Process
When the trainer is started by the Kubernetes, it executes the following steps at startup:
1. Watches the available parameter server prefix keys `/ps/` on etcd and waits until the count of parameter servers reaches the desired count.
1. Generates a unique ID, and sets key `/trainer/<unique ID>` with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.
1. Waits for tasks from the master to start training.
If trainer's etcd lease expires, it will try set key `/trainer/<unique ID>` again so that the master process can discover the trainer again.
### Parameter Server Process
When the parameter server is started by Kubernetes, it executes the following steps at startup:
1. Read desired total number of parameter servers from etcd `/ps_desired`
1. Search through etcd keys `/ps/<index>` (`/ps/0`, `/ps/1`, ...) to find the first non-existant key whose index is smaller than the total number of parameter servers. Set the key using a transaction to avoid concurrent writes. The parameter server's index is inferred from the key name.
The desired number of parameter servers is 3:
<img src="src/paddle-ps-0.png"/>
The third parameter server joined:
<img src="src/paddle-ps-1.png"/>
1. The parameter server can load parameters if there are already saved parameters in the save path (inferred from its index).
1. Now the parameter server is ready for the trainers' requests.
If the parameter server's etcd lease expires, the parameter server will kill itself.
<p>In <aclass="reference external"href="https://www.slideshare.net/cxwangyi/paddlepaddle-a-complete-solution-for-businesses">this slides</a>, we explained that we’d like PaddlePaddle running on general-purpose clusters like those managed by Kubernetes, so to address demands for AI from both Internet and non-Internet industries.</p>
<p>This poses technical challenges to PaddlePaddle:</p>
<olclass="simple">
<li>Support fault-recovery.</li>
<li>Support both offline and online training.</li>
<li><aclass="reference external"href="https://en.wikipedia.org/wiki/Serverless_computing">Serverless computing</a> of distributed training.</li>
<p>A training job will be created once user asks Paddle cloud to train a model. The training job is made up of different processes that collaboratively consume data and produce a trained model. There are three kinds of processes:</p>
<olclass="simple">
<li>the <em>master process</em>, which dispatches tasks to</li>
<li>one or more <em>trainer processes</em>, which run distributed training and synchronize gradients/models via</li>
<li>one or more <em>parameter server processes</em>, where each holds a shard of the global model.</li>
</ol>
<p>Their relation is illustrated in the following graph:</p>
<li>Partition a dataset into <aclass="reference external"href="#task">tasks</a> and dispatch tasks to trainers.</li>
<li>Keep track of training progress on the dataset with <aclass="reference external"href="#task-queue">task queue</a>. A training job will iterate on the dataset for a full pass until it goes into next pass.</li>
<p>A task is a data shard to be trained. The total number of tasks will be much bigger than the total number of trainers. The number of data instances inside a task will be much bigger than the mini-batch size.</p>
<p>The master process has three task queues to track training progress. As illustrated in the graph below, Job A and Job B both have one master process. Each master process has three task queues.</p>
<p><imgsrc="src/paddle-task-queues.png"/></p>
<ulclass="simple">
<li>The todo queue holds tasks to be dispatched. When a job starts, the master process fills in the todo queue with all tasks.</li>
<li>The pending queue holds tasks that are currently training by trainers.</li>
<li>the done queue holds tasks that are already trained.</li>
</ul>
<p>The life cycle of a single task is illustrated below:</p>
<p><imgsrc="src/paddle-task-states.png"/></p>
<olclass="simple">
<li>When a new pass of training starts, all tasks will be placed in the todo queue.</li>
<li>The master process will dispatch few tasks to each trainer at a time, puts them in the pending queue and waits for completion.</li>
<li>The trainer will work on its tasks and tell the master process once a task is completed. The master process will dispatch a new task to that trainer.</li>
<li>If a task timeout. the master process will move it back to the todo queue. The timeout count will increase by one. If the timeout count is above a threshold, the task is likely to cause a trainer to crash, so it will be discarded.</li>
<li>The master process will move completed task to the done queue. When the todo queue is empty, the master process will start a new pass by moving all tasks in the done queue to todo queue and reset the timeout counter of all tasks to zero.</li>
<li>Work on the tasks: calculate and upload gradient to parameter servers, and update local model by downloading new parameters from parameter servers.</li>
</ul>
</div>
<divclass="section"id="parameter-server-process">
<spanid="parameter-server-process"></span><h3>Parameter Server Process<aclass="headerlink"href="#parameter-server-process"title="永久链接至标题">¶</a></h3>
<p>Parameter server processes hold the parameters collaboratively. The parameters are partitioned on different parameter servers.</p>
<p>The parameter server will:</p>
<ulclass="simple">
<li>Receive gradient from the trainers, update its parameters, and give the trainers the latest parameters.</li>
<li>Periodically save its parameters to distributed file system by overriding the previous save.</li>
<p>Parameter server will wait for all trainer finish n-th mini-batch calculation and send their gradients before broadcasting new parameters to every trainer. Every trainer will wait for the new parameters before starting n+1-th mini-batch.</p>
<p>The training job will pause if the master processes is dead, or any of the parameter server process is dead. They will be started by <aclass="reference external"href="https://kubernetes.io/">Kubernetes</a> and recover in few minutes. Please refer to <aclass="reference external"href="#fault-recovery">fault recovery</a>.</p>
<p>The training job will continue to make progress if there is at least one training process running. The strategy depends on the type of optimization algorithm:</p>
<ul>
<li><pclass="first">sync-SGD</p>
<p>TODO</p>
</li>
<li><pclass="first">async-SGD</p>
<p>Since async-SGD does not require synchronization between mini-batches, the system will by definition make process if at least one trainer is running.</p>
<p>PaddlePaddle uses <aclass="reference external"href="https://github.com/coreos/etcd">etcd</a> to keep track of the states of processes. Because etcd is a distributed reliable key-value store, the restarted process can recover its states from etcd. The model parameters are periodically saved into distributed file system, so a restarted parameter server can recover its parameters from the saved file.</p>
<p>Now we will introduce how each process recovers from a failure, the graph below shows how etcd is used:</p>
<p>When the master is started by the Kubernetes, it executes the following steps at startup:</p>
<olclass="simple">
<li>Grabs a unique <em>master</em> lock in etcd, which prevents concurrent master instantiations.</li>
<li>Recovers the task queues from etcd if they already exist, otherwise, the master will create them.</li>
<li>Watches the trainer prefix keys <codeclass="docutils literal"><spanclass="pre">/trainer/</span></code> on etcd to find the live trainers.</li>
<li>Starts dispatching the tasks to the trainers, and updates task queue using an etcd transaction to ensure lock is held during the update.</li>
</ol>
<p>The master process will kill itself if its etcd lease expires.</p>
<p>When the master process is dead for any reason, Kubernetes will restart it. It will be online again with all states recovered from etcd in few minutes.</p>
<p>When the trainer is started by the Kubernetes, it executes the following steps at startup:</p>
<olclass="simple">
<li>Watches the available parameter server prefix keys <codeclass="docutils literal"><spanclass="pre">/ps/</span></code> on etcd and waits until the count of parameter servers reaches the desired count.</li>
<li>Generates a unique ID, and sets key <codeclass="docutils literal"><spanclass="pre">/trainer/<unique</span><spanclass="pre">ID></span></code> with its contact address as value. The key will be deleted when the lease expires, so the master will be aware of the trainer being online and offline.</li>
<li>Waits for tasks from the master to start training.</li>
</ol>
<p>If trainer’s etcd lease expires, it will try set key <codeclass="docutils literal"><spanclass="pre">/trainer/<unique</span><spanclass="pre">ID></span></code> again so that the master process can discover the trainer again.</p>
</div>
<divclass="section"id="parameter-server-process">
<spanid="id3"></span><h3>Parameter Server Process<aclass="headerlink"href="#parameter-server-process"title="永久链接至标题">¶</a></h3>
<p>When the parameter server is started by Kubernetes, it executes the following steps at startup:</p>
<ol>
<li><pclass="first">Read desired total number of parameter servers from etcd <codeclass="docutils literal"><spanclass="pre">/ps_desired</span></code></p>
</li>
<li><pclass="first">Search through etcd keys <codeclass="docutils literal"><spanclass="pre">/ps/<index></span></code> (<codeclass="docutils literal"><spanclass="pre">/ps/0</span></code>, <codeclass="docutils literal"><spanclass="pre">/ps/1</span></code>, ...) to find the first non-existant key whose index is smaller than the total number of parameter servers. Set the key using a transaction to avoid concurrent writes. The parameter server’s index is inferred from the key name.</p>
<p>The desired number of parameter servers is 3:</p>
<p><imgsrc="src/paddle-ps-0.png"/></p>
<p>The third parameter server joined:</p>
<p><imgsrc="src/paddle-ps-1.png"/></p>
</li>
<li><pclass="first">The parameter server can load parameters if there are already saved parameters in the save path (inferred from its index).</p>
</li>
<li><pclass="first">Now the parameter server is ready for the trainers’ requests.</p>
</li>
</ol>
<p>If the parameter server’s etcd lease expires, the parameter server will kill itself.</p>
Built with <ahref="http://sphinx-doc.org/">Sphinx</a> using a <ahref="https://github.com/snide/sphinx_rtd_theme">theme</a> provided by <ahref="https://readthedocs.org">Read the Docs</a>.
