In this article, we explain how to run distributed Paddle training jobs on clusters. We will create the distributed version of the single-process training example, [recommendation](https://github.com/baidu/Paddle/tree/develop/demo/recommendation).
[Scripts](https://github.com/baidu/Paddle/tree/develop/paddle/scripts/cluster_train) used in this article launch distributed jobs via SSH. They also work as a reference for users running more sophisticated cluster management systems like MPI and Kubernetes.
[Scripts](https://github.com/baidu/Paddle/tree/develop/paddle/scripts/cluster_train) used in this article launch distributed jobs via SSH. They also work as a reference for users running more sophisticated cluster management systems like MPI and [Kubernetes](https://github.com/PaddlePaddle/Paddle/tree/develop/doc/howto/usage/cluster/k8s).
## Prerequisite
...
...
@@ -20,13 +20,13 @@ In this article, we explain how to run distributed Paddle training jobs on clust
We refer to the directory where we put dependent libraries, config files, etc., as *workspace*.
These ```train/test``` data should be prepared before launching cluster job. To satisfy the requirement that train/test data are placed in different directory from workspace, PADDLE refers train/test data according to index file named as ```train.list/test.list``` which are used in model config file. So the train/test data also contains train.list/test.list two list file. All local training demo already provides scripts to help you create these two files, and all nodes in cluster job will handle files with same logical code in normal condition.
These `train/test` data should be prepared before launching cluster job. To satisfy the requirement that train/test data are placed in different directory from workspace, PADDLE refers train/test data according to index file named as `train.list/test.list` which are used in model config file. So the train/test data also contains train.list/test.list two list file. All local training demo already provides scripts to help you create these two files, and all nodes in cluster job will handle files with same logical code in normal condition.
Generally, you can use same model file from local training for cluster training. What you should have in mind that, the ```batch_size``` set in ```setting``` function in model file means batch size in ```each``` node of cluster job instead of total batch size if synchronization SGD was used.
Generally, you can use same model file from local training for cluster training. What you should have in mind that, the `batch_size` set in `setting` function in model file means batch size in `each` node of cluster job instead of total batch size if synchronization SGD was used.
Following steps are based on demo/recommendation demo in demo directory.
Following steps are based on [demo/recommendation](https://github.com/PaddlePaddle/Paddle/tree/develop/demo/recommendation) demo in demo directory.
You just go through demo/recommendation tutorial doc until ```Train``` section, and at last you will get train/test data and model configuration file. Finaly, just use demo/recommendation as workspace for cluster training.
You just go through demo/recommendation tutorial doc until `Train` section, and at last you will get train/test data and model configuration file. Finaly, just use demo/recommendation as workspace for cluster training.
At last your workspace should look like as follow:
```
...
...
@@ -55,16 +55,16 @@ At last your workspace should look like as follow:
```
Not all of these files are needed for cluster training, but it's not necessary to remove useless files.
```trainer_config.py```
`trainer_config.py`
Indicates the model config file.
```train.list``` and ```test.list```
`train.list` and `test.list`
File index. It stores all relative or absolute file paths of all train/test data at current node.
```dataprovider.py```
`dataprovider.py`
used to read train/test samples. It's same as local training.
```data```
`data`
all files in data directory are refered by train.list/test.list which are refered by data provider.
...
...
@@ -72,19 +72,19 @@ all files in data directory are refered by train.list/test.list which are refere
The options below must be carefully set in cluster_train/conf.py
```HOSTS``` all nodes hostname or ip that will run cluster job. You can also append user and ssh port with hostname, such as root@192.168.100.17:9090.
`HOSTS` all nodes hostname or ip that will run cluster job. You can also append user and ssh port with hostname, such as root@192.168.100.17:9090.
```ROOT_DIR``` workspace ROOT directory for placing JOB workspace directory
`ROOT_DIR` workspace ROOT directory for placing JOB workspace directory
```PADDLE_NIC``` the NIC(Network Interface Card) interface name for cluster communication channel, such as eth0 for ethternet, ib0 for infiniband.
`PADDLE_NIC` the NIC(Network Interface Card) interface name for cluster communication channel, such as eth0 for ethternet, ib0 for infiniband.
```PADDLE_PORT``` port number for cluster commnunication channel
`PADDLE_PORT` port number for cluster commnunication channel
```PADDLE_PORTS_NUM``` the number of port used for cluster communication channle. if the number of cluster nodes is small(less than 5~6nodes), recommend you set it to larger, such as 2 ~ 8, for better network performance.
`PADDLE_PORTS_NUM` the number of port used for cluster communication channle. if the number of cluster nodes is small(less than 5~6nodes), recommend you set it to larger, such as 2 ~ 8, for better network performance.
```PADDLE_PORTS_NUM_FOR_SPARSE``` the number of port used for sparse updater cluster commnunication channel. if sparse remote update is used, set it like ```PADDLE_PORTS_NUM```
`PADDLE_PORTS_NUM_FOR_SPARSE` the number of port used for sparse updater cluster commnunication channel. if sparse remote update is used, set it like `PADDLE_PORTS_NUM`
```LD_LIBRARY_PATH``` set addtional LD_LIBRARY_PATH for cluster job. You can use it to set CUDA libraries path.
`LD_LIBRARY_PATH` set addtional LD_LIBRARY_PATH for cluster job. You can use it to set CUDA libraries path.
```paddle.py``` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can set as ```paddle.py``` command options and ```paddle.py``` will transparently and automatically set these options to PaddlePaddle lower level processes.
`paddle.py` provides automatical scripts to start all PaddlePaddle cluster processes in different nodes. By default, all command line options can set as `paddle.py` command options and `paddle.py` will transparently and automatically set these options to PaddlePaddle lower level processes.
```paddle.py```provides two distinguished command option for easy job launching.
`paddle.py`provides two distinguished command option for easy job launching.
```job_dispatch_package``` set it with local ```workspace```directory, it will be dispatched to all nodes set in conf.py. It could be helpful for frequent hacking workspace files, otherwise frequent mulit-nodes workspace deployment could make your crazy.
```job_workspace``` set it with already deployed workspace directory, ```paddle.py``` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy
`job_dispatch_package` set it with local `workspace`directory, it will be dispatched to all nodes set in conf.py. It could be helpful for frequent hacking workspace files, otherwise frequent mulit-nodes workspace deployment could make your crazy.
`job_workspace` set it with already deployed workspace directory, `paddle.py` will skip dispatch stage to directly launch cluster job with all nodes. It could help to reduce heavy
dispatch latency.
```cluster_train/run.sh``` provides command line sample to run ```demo/recommendation``` cluster job, just modify ```job_dispatch_package``` and ```job_workspace``` with your defined directory, then:
`cluster_train/run.sh` provides command line sample to run `demo/recommendation` cluster job, just modify `job_dispatch_package` and `job_workspace` with your defined directory, then:
```
sh run.sh
```
...
...
@@ -134,23 +134,23 @@ sh run.sh
The cluster Job will start in several seconds.
### Kill Cluster Job
```paddle.py``` can capture ```Ctrl + C``` SIGINT signal to automatically kill all processes launched by it. So just stop ```paddle.py``` to kill cluster job. You should mannally kill job if program crashed.
`paddle.py` can capture `Ctrl + C` SIGINT signal to automatically kill all processes launched by it. So just stop `paddle.py` to kill cluster job. You should mannally kill job if program crashed.
### Check Cluster Training Result
Check log in $workspace/log for details, each node owns same log structure.
```paddle_trainer.INFO```
`paddle_trainer.INFO`
It provides almost all interal output log for training, same as local training. Check runtime model convergence here.
```paddle_pserver2.INFO```
`paddle_pserver2.INFO`
It provides pserver running log, which could help to diagnose distributed error.
```server.log```
`server.log`
It provides stderr and stdout of pserver process. Check error log if training crashs.
```train.log```
`train.log`
It provides stderr and stdout of trainer process. Check error log if training crashs.
### Check Model Output
After one pass finished, model files will be writed in ```output``` directory in node 0.
```nodefile``` in workspace indicates the node id of current cluster job.
After one pass finished, model files will be writed in `output` directory in node 0.
`nodefile` in workspace indicates the node id of current cluster job.
You can refer to the following link for more detailed examples and data formats: <ahref = "../../api/data_provider/pydataprovider2_en.html">PyDataProvider2</a>.
## Network Architecture
You will describe four kinds of network architectures in this section.
We will describe four kinds of network architectures in this section.
First, you will build a logistic regression model. Later, you will also get chance to build other more powerful network architectures.
...
...
@@ -391,7 +391,7 @@ paddle train \
--use_gpu=false
```
We do not provide examples on how to train on clusters here. If you want to train on clusters, please follow the <ahref = "../../howto/cluster/cluster_train_en.html">distributed training</a> documentation or other demos for more details.
We do not provide examples on how to train on clusters here. If you want to train on clusters, please follow the <ahref = "../../howto/usage/cluster/cluster_train_en.html">distributed training</a> documentation or other demos for more details.
## Inference
You can use the trained model to perform prediction on the dataset with no labels. You can also evaluate the model on dataset with labels to obtain its test accuracy.
...
...
@@ -509,7 +509,7 @@ The scripts of data downloading, network configurations, and training scrips are
*\--config_args：Other configuration arguments.
*\--init_model_path：The path of the initial model parameter.
By default, the trainer will save model every pass. You can also specify `saving_period_by_batches` to set the frequency of batch saving. You can use `show_parameter_stats_period` to print the statistics of the parameters, which are very useful for tuning parameters. Other command line arguments can be found in <ahref = "../../howto/cmd_parameter/index_en.html">command line argument documentation</a>。
By default, the trainer will save model every pass. You can also specify `saving_period_by_batches` to set the frequency of batch saving. You can use `show_parameter_stats_period` to print the statistics of the parameters, which are very useful for tuning parameters. Other command line arguments can be found in <ahref = "../../howto/usage/cmd_parameter/index_en.html">command line argument documentation</a>。
