!438 Add parameter server tutorial

Merge pull request !438 from ZPaC/add-ps

tutorials/source_zh_cn/advanced_use/distributed_training_tutorials.rst

@@ -6,4 +6,5 @@
 
   distributed_training
   host_device_training
-  checkpoint_for_hybrid_parallel
+  checkpoint_for_hybrid_parallel
\ No newline at end of file
+  parameter_server_training
\ No newline at end of file

tutorials/source_zh_cn/advanced_use/parameter_server_training.md

+# Parameter Server训练
+
+<!-- TOC -->
+
+- [Parameter Server训练](#parameter_server训练)
+    - [概述](#概述)
+    - [准备工作](#准备工作)
+        - [训练脚本准备](#训练脚本准备)
+        - [参数设置](#参数设置)
+        - [环境变量设置](#环境变量设置)
+    - [执行训练](#执行训练)
+
+<!-- /TOC -->
+
+<a href="https://gitee.com/mindspore/docs/blob/master/tutorials/source_zh_cn/advanced_use/parameter_server_training.md" target="_blank"><img src="../_static/logo_source.png"></a>
+
+## 概述
+
+## 准备工作
+以LeNet在Ascend 910上使用Parameter Server，并且配置单Worker，单Server训练为例：
+
+### 训练脚本准备
+
+参考<https://gitee.com/mindspore/mindspore/tree/master/model_zoo/official/cv/lenet>，了解如何训练一个LeNet网络。
+
+### 参数设置
+
+在本训练模式下，有以下两种调用接口方式以控制训练参数是否通过Parameter Server进行更新：
+  
+- 通过`mindspore.nn.Cell.set_param_ps()`对`nn.Cell`中所有权重递归设置
+- 通过`mindspore.common.Parameter.set_param_ps()`对此权重进行设置
+
+在原训练脚本基础上，设置LeNet模型所有权重通过Parameter Server训练：
+```python
+network = LeNet5(cfg.num_classes)
+network.set_param_ps()
+```
+
+### 环境变量设置
+
+Mindspore通过读取环境变量，控制Parameter Server训练，环境变量包括以下选项：
+
+```
+export MS_SERVER_NUM=1                # Server number
+export MS_WORKER_NUM=1                # Worker number
+export MS_SCHED_HOST=XXX.XXX.XXX.XXX  # Scheduler IP address
+export MS_SCHED_POST=XXXX             # Scheduler port
+export MS_ROLE=MS_SCHED               # The role of this process: MS_SCHED represents the scheduler, MS_WORKER represents the worker, MS_PSERVER represents the Server
+```
+
+## 执行训练
+
+1. shell脚本
+
+    提供Worker，Server和Scheduler三个角色对应的shell脚本，以启动训练：
+
+    `Scheduler.sh`:
+    ```bash
+    #!/bin/bash
+    export MS_SERVER_NUM=1
+    export MS_WORKER_NUM=1
+    export MS_SCHED_HOST=XXX.XXX.XXX.XXX
+    export MS_SCHED_POST=XXXX
+    export MS_ROLE=MS_SCHED
+    python train.py
+    ```
+
+    `Server.sh`:
+    ```bash
+    #!/bin/bash
+    export MS_SERVER_NUM=1
+    export MS_WORKER_NUM=1
+    export MS_SCHED_HOST=XXX.XXX.XXX.XXX
+    export MS_SCHED_POST=XXXX
+    export MS_ROLE=MS_PSERVER
+    python train.py
+    ```
+
+    `Worker.sh`:
+    ```bash
+    #!/bin/bash
+    export MS_SERVER_NUM=1
+    export MS_WORKER_NUM=1
+    export MS_SCHED_HOST=XXX.XXX.XXX.XXX
+    export MS_SCHED_POST=XXXX
+    export MS_ROLE=MS_WORKER
+    python train.py
+    ```
+
+    最后分别执行：
+    ```bash
+    sh Scheduler.sh > scheduler.log 2>&1 &
+    sh Server.sh > server.log 2>&1 &
+    sh Worker.sh > worker.log 2>&1 &
+    ```
+    启动训练
+
+2. 查看结果
+
+    查看`scheduler.log`中和Server与Worker通信日志：
+    ```
+    Bind to role=scheduler, id=1, ip=XXX.XXX.XXX.XXX, port=XXXX
+    Assign rank=8 to node role=server, ip=XXX.XXX.XXX.XXX, port=XXXX
+    Assign rank=9 to node role=worker, ip=XXX.XXX.XXX.XXX, port=XXXX
+    the scheduler is connected to 1 workers and 1 servers
+    ```
+    说明Server、Worker与Scheduler通信建立成功。
+
+    查看`worker.log`中训练结果：
+    ```
+    epoch: 1 step: 1, loss is 2.302287
+    epoch: 1 step: 2, loss is 2.304071
+    epoch: 1 step: 3, loss is 2.308778
+    epoch: 1 step: 4, loss is 2.301943
+    ...
+    ```