# Paddle on Kubernetes:分布式训练 前一篇文章介绍了如何在Kubernetes集群上启动一个单机Paddle训练作业 (Job)。在这篇文章里,我们介绍如何在Kubernetes集群上启动分布式Paddle训练作业。关于Paddle的分布式集群训练,可以参考 [Cluster Training](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md), 本文在此基础上,利用Kubernetes快速构建Paddle集群,进行分布式训练任务。 ## 制作镜像 Paddle的集群训练需要有一个Paddle集群来实现,在本文中,我们使用Kubernetes来快速创建一个Paddle集群。我们使用 `paddledev/paddle:cpu-demo-latest` 镜像作为Paddle集群节点的运行环境,里面包含了 Paddle 运行所需要的相关依赖,同时,为了能将训练任务及配置统一分发到各个节点,需要使用到`sshd`以便使用`fabric`来操作。镜像的 Dockerfile 如下: ``` FROM paddledev/paddle:cpu-demo-latest RUN apt-get update RUN apt-get install -y openssh-server RUN mkdir /var/run/sshd RUN echo 'root:root' | chpasswd RUN sed -ri 's/^PermitRootLogin\s+.*/PermitRootLogin yes/' /etc/ssh/sshd_config RUN sed -ri 's/UsePAM yes/#UsePAM yes/g' /etc/ssh/sshd_config EXPOSE 22 CMD ["/usr/sbin/sshd", "-D"] ``` 使用 `docker build` 构建镜像: ``` docker build -t mypaddle:paddle_demo_ssh . ``` ## 准备工作空间 工作空间 [Job Workspace](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md#prepare-job-workspace) , 即一个包含了依赖库,训练,测试数据,模型配置文件的目录。参考 [Cluster Training](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md)中的例子,我们也是用`demo/recommendation`作为本文的训练任务。此demo可直接从[Github Paddle源码](https://github.com/baidu/Paddle/tree/develop/demo/recommendation)中获取。 ### 准备训练数据 在Paddle源码中,找到`demo/recommendation`文件夹,即为我们的Workspace, 在本文的环境中,路径为`/home/work/paddle-demo/Paddle/demo/recommendation` ``` [root@paddle-k8s-node0 recommendation]# tree . ├── common_utils.py ├── data │   ├── config_generator.py │   ├── config.json │   ├── meta_config.json │   ├── meta_generator.py │   ├── ml_data.sh │   └── split.py ├── dataprovider.py ├── evaluate.sh ├── prediction.py ├── preprocess.sh ├── requirements.txt ├── run.sh └── trainer_config.py 1 directory, 14 files ``` 运行`data/ml_data.sh`脚本,下载数据,然后运行`preprocess.sh`脚本进行预处理。 ``` [root@paddle-k8s-node0 recommendation]# data/ml_data.sh ++ dirname data/ml_data.sh + cd data + wget http://files.grouplens.org/datasets/movielens/ml-1m.zip --2016-11-04 10:14:49-- http://files.grouplens.org/datasets/movielens/ml-1m.zip Resolving files.grouplens.org (files.grouplens.org)... 128.101.34.146 Connecting to files.grouplens.org (files.grouplens.org)|128.101.34.146|:80... connected. HTTP request sent, awaiting response... 200 OK Length: 5917549 (5.6M) [application/zip] Saving to: ‘ml-1m.zip’ 100%[==========================>] 5,917,549 50.6KB/s in 2m 29s 2016-11-04 10:17:20 (38.8 KB/s) - ‘ml-1m.zip’ saved [5917549/5917549] + unzip ml-1m.zip Archive: ml-1m.zip creating: ml-1m/ inflating: ml-1m/movies.dat inflating: ml-1m/ratings.dat inflating: ml-1m/README inflating: ml-1m/users.dat + rm ml-1m.zip [root@paddle-k8s-node0 recommendation]# ./preprocess.sh generate meta config file generate meta file split train/test file shuffle train file ``` ### 修改集群训练配置 参考[Cluster Training](https://github.com/baidu/Paddle/blob/develop/doc/cluster/opensource/cluster_train.md)中的介绍,我们使用`paddle/scripts/cluster_train/`中的文件来作为分布式训练任务的配置和启动脚本。在`run.sh`文件中,填入我们的workspace和训练配置文件路径。 ``` #!/bin/sh python paddle.py \ --job_dispatch_package="/home/work/paddle-demo/Paddle/demo/recommendation" \ --dot_period=10 \ --ports_num_for_sparse=2 \ --log_period=50 \ --num_passes=10 \ --trainer_count=4 \ --saving_period=1 \ --local=0 \ --config=/home/work/paddle-demo/Paddle/demo/recommendation/trainer_config.py \ --save_dir=./output \ --use_gpu=0 ``` ## 创建Paddle集群 创建Paddle集训需要编写创建Kubernetes资源的yaml文件,首先,创建一个Service,便于我们通过此Service来查找其对应的Paddle节点。 ``` apiVersion: v1 kind: Service metadata: name: cluster-demo spec: selector: app: cluster-demo ports: - name: default protocol: TCP port: 7164 targetPort: 7164 ``` 为了创建多个Paddle节点,我们使用Kubernetes ReplicationController资源来控制Paddle集群中的节点数量,Paddle节点之间需要开放相关的端口来互相通信。下面的例子中,我们开放了每个Paddle节点的7164-7167端口,例如,一个包含4个节点的Paddle集群的yaml文件如下: ``` apiVersion: v1 kind: ReplicationController metadata: name: cluster-demo spec: replicas: 4 selector: app: cluster-demo template: metadata: name: cluster-demo labels: app: cluster-demo spec: containers: - name: cluster-demo image: mypaddle:paddle_demo_ssh ports: - containerPort: 7164 - containerPort: 7165 - containerPort: 7166 - containerPort: 7167 ``` 然后我们可以通过`kubectl`工具来查看所创建的资源信息。 首先查看我们创建的Paddle Service,然后根据Service,查看所创建的Paddle节点的IP地址。 ``` [root@paddle-k8s-node0 cluster_train]# kubectl get svc NAME CLUSTER-IP EXTERNAL-IP PORT(S) AGE cluster-demo 11.1.1.77 7164/TCP 6h [root@paddle-k8s-node0 cluster_train]# kubectl get -o json endpoints cluster-demo | grep ip "ip": "192.168.129.79", "ip": "192.168.129.80", "ip": "192.168.223.157", "ip": "192.168.223.158", ``` ## 开始集群训练 我们需要在`paddle/scripts/cluster_train/conf.py`文件中指定各个节点的IP地址以及开放的端口。根据上文创建的信息,`conf.py`文件修改如下: ``` HOSTS = [ "root@192.168.129.79", "root@192.168.129.80", "root@192.168.223.157", "root@192.168.223.158" ] ''' workspace configuration ''' #root dir for workspace, can be set as any director with real user account ROOT_DIR = "/home/paddle" ''' network configuration ''' #pserver nics PADDLE_NIC = "eth0" #pserver port PADDLE_PORT = 7164 #pserver ports num PADDLE_PORTS_NUM = 2 #pserver sparse ports num PADDLE_PORTS_NUM_FOR_SPARSE = 2 #environments setting for all processes in cluster job LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/lib64" ``` 然后使用`run.sh`脚本开始训练,启动的打印如下: ``` [root@paddle-k8s-node0 cluster_train]# ./run.sh [root@192.168.129.79] Executing task 'job_create_workspace' ...... [root@192.168.129.80] Executing task 'job_create_workspace' ...... [root@192.168.223.157] Executing task 'job_create_workspace' ...... [root@192.168.223.158] Executing task 'job_create_workspace' ...... [root@192.168.129.79] run: echo 0 > /home/paddle/JOB20161104171630/nodefile [root@192.168.129.80] Executing task 'set_nodefile' [root@192.168.129.80] run: echo 1 > /home/paddle/JOB20161104171630/nodefile [root@192.168.223.157] Executing task 'set_nodefile' [root@192.168.223.157] run: echo 2 > /home/paddle/JOB20161104171630/nodefile [root@192.168.223.158] Executing task 'set_nodefile' [root@192.168.223.158] run: echo 3 > /home/paddle/JOB20161104171630/nodefile ``` 可以看到192.168.129.79,192.168.129.80,192.168.223.157,192.168.223.158分别为Paddle集群的Node 0-3. 我们可以进入其中一个Paddle节点查看训练的日志。 ``` root@cluster-demo-fwwi5:/home/paddle/JOB20161104171700/log# less paddle_trainer.INFO Log file created at: 2016/11/04 09:17:20 Running on machine: cluster-demo-fwwi5 Log line format: [IWEF]mmdd hh:mm:ss.uuuuuu threadid file:line] msg I1104 09:17:20.346797 108 Util.cpp:155] commandline: /usr/local/bin/../opt/paddle/bin/paddle _trainer --num_gradient_servers=4 --nics=eth0 --port=7164 --ports_num=2 --comment=paddle_proce ss_by_paddle --pservers=192.168.129.79,192.168.129.80,192.168.223.157,192.168.223.158 --ports_ num_for_sparse=2 --config=./trainer_config.py --trainer_count=4 --use_gpu=0 --num_passes=10 -- save_dir=./output --log_period=50 --dot_period=10 --saving_period=1 --local=0 --trainer_id=1 root@cluster-demo-fwwi5:/home/paddle/JOB20161104171700/log# tailf paddle_trainer.INFO ...... I1104 09:17:37.376471 150 ThreadLocal.cpp:37] thread use undeterministic rand seed:151 I1104 09:18:54.159624 108 TrainerInternal.cpp:163] Batch=50 samples=80000 AvgCost=4.03478 CurrentCost=4.03478 Eval: CurrentEval: I1104 09:20:10.207902 108 TrainerInternal.cpp:163] Batch=100 samples=160000 AvgCost=3.75806 CurrentCost=3.48134 Eval: CurrentEval: I1104 09:21:26.493571 108 TrainerInternal.cpp:163] Batch=150 samples=240000 AvgCost=3.64512 CurrentCost=3.41923 Eval: CurrentEval: ``` 最后,我们可以在Paddle集群的node0(192.168.129.79)上查看训练的输出结果。 ``` [root@paddle-k8s-node0 ~]# ssh root@192.168.129.79 ...... root@cluster-demo-r65g0:/home/paddle/JOB20161104171700/output/pass-00000# ll total 14876 drwxr-xr-x. 2 root root 4096 Nov 4 09:40 ./ drwxr-xr-x. 3 root root 23 Nov 4 09:40 ../ -rw-r--r--. 1 root root 4046864 Nov 4 09:40 ___embedding_0__.w0 -rw-r--r--. 1 root root 100368 Nov 4 09:40 ___embedding_1__.w0 -rw-r--r--. 1 root root 6184976 Nov 4 09:40 ___embedding_2__.w0 -rw-r--r--. 1 root root 2064 Nov 4 09:40 ___embedding_3__.w0 -rw-r--r--. 1 root root 7184 Nov 4 09:40 ___embedding_4__.w0 -rw-r--r--. 1 root root 21520 Nov 4 09:40 ___embedding_5__.w0 -rw-r--r--. 1 root root 262160 Nov 4 09:40 ___fc_layer_0__.w0 -rw-r--r--. 1 root root 1040 Nov 4 09:40 ___fc_layer_0__.wbias ...... ...... -rw-r--r--. 1 root root 262160 Nov 4 09:40 _movie_fusion.w0 -rw-r--r--. 1 root root 262160 Nov 4 09:40 _movie_fusion.w1 -rw-r--r--. 1 root root 262160 Nov 4 09:40 _movie_fusion.w2 -rw-r--r--. 1 root root 1040 Nov 4 09:40 _movie_fusion.wbias -rw-r--r--. 1 root root 262160 Nov 4 09:40 _user_fusion.w0 -rw-r--r--. 1 root root 262160 Nov 4 09:40 _user_fusion.w1 -rw-r--r--. 1 root root 262160 Nov 4 09:40 _user_fusion.w2 -rw-r--r--. 1 root root 262160 Nov 4 09:40 _user_fusion.w3 -rw-r--r--. 1 root root 1040 Nov 4 09:40 _user_fusion.wbias -rw-r--r--. 1 root root 169 Nov 4 09:40 done -rw-r--r--. 1 root root 17 Nov 4 09:40 path.txt -rw-r--r--. 1 root root 3495 Nov 4 09:40 trainer_config.py ```