ipex-llm/docs/readthedocs/source/doc/UserGuide/hadoop.md

Hadoop/YARN User Guide

Hadoop version: Apache Hadoop >= 2.7 (3.X included) or CDH 5.X. CDH 6.X have not been tested and thus currently not supported.

---

For Scala users, please see Scala User Guide for how to run BigDL on Hadoop/YARN clusters.

For Python users, you can run BigDL programs on standard Hadoop/YARN clusters without any changes to the cluster (i.e., no need to pre-install BigDL or other Python libraries on all nodes in the cluster).

1. Prepare Python Environment

• You need to first use conda to prepare the Python environment on the local machine where you submit your application. Create a conda environment, install BigDL and all the needed Python libraries in the created conda environment:

  conda create -n bigdl python=3.7  # "bigdl" is conda environment name, you can use any name you like.
  conda activate bigdl
  
  pip install bigdl
  
  # Use conda or pip to install all the needed Python dependencies in the created conda environment.
  

  View the Python User Guide for more details for BigDL installation.

• You need to download and install JDK in the environment, and properly set the environment variable JAVA_HOME, which is required by Spark. JDK8 is highly recommended.

  You may take the following commands as a reference for installing OpenJDK:

  # For Ubuntu
  sudo apt-get install openjdk-8-jre
  export JAVA_HOME=/usr/lib/jvm/java-8-openjdk-amd64/
  
  # For CentOS
  su -c "yum install java-1.8.0-openjdk"
  export JAVA_HOME=/usr/lib/jvm/java-1.8.0-openjdk-1.8.0.282.b08-1.el7_9.x86_64/jre
  
  export PATH=$PATH:$JAVA_HOME/bin
  java -version  # Verify the version of JDK.
  

• Check the Hadoop setup and configurations of your cluster. Make sure you properly set the environment variable HADOOP_CONF_DIR, which is needed to initialize Spark on YARN:

  export HADOOP_CONF_DIR=the directory of the hadoop and yarn configurations
  

• For CDH users

  If your CDH cluster has already installed Spark, the CDH's Spark might be conflict with the pyspark installed by pip required by BigDL.

  Thus before running BigDL applications, you should unset all the Spark related environment variables. You can use env | grep SPARK to find all the existing Spark environment variables.

  Also, a CDH cluster's HADOOP_CONF_DIR should be /etc/hadoop/conf on CDH by default.

---

2. Run on YARN with built-in function

This is the easiest and most recommended way to run BigDL on YARN, as you don't need to care about environment preparation and Spark related commands. In this way, you can easily switch your job between local (for test) and YARN (for production) by changing the "cluster_mode".

• Call init_orca_context at the very beginning of your code to initiate and run BigDL on standard Hadoop/YARN clusters:

  from bigdl.orca import init_orca_context
  
  sc = init_orca_context(cluster_mode="yarn-client", cores=4, memory="10g", num_nodes=2)
  

  init_orca_context would automatically prepare the runtime Python environment, detect the current Hadoop configurations from HADOOP_CONF_DIR and initiate the distributed execution engine on the underlying YARN cluster. View Orca Context for more details.

  By specifying "cluster_mode" to be yarn-client or yarn-cluster, init_orca_context will submit the job to YARN with client and cluster mode respectively.

  The difference between yarn-client and yarn-cluster is where you run your Spark driver. For yarn-client, the Spark driver will run on the node where you start Python, while for yarn-cluster the Spark driver will run on a random node in the YARN cluster. So if you are running with yarn-cluster, you should change the application's data loading from local file to a network file system (e.g. HDFS).

• You can then simply run your BigDL program in a Jupyter notebook. Note that jupyter cannot run on yarn-cluster, as the driver is not running on the local node.

  jupyter notebook --notebook-dir=./ --ip=* --no-browser
  

  Or you can run your BigDL program as a normal Python script (e.g. script.py) and in this case both yarn-client and yarn-cluster are supported.

  python script.py
  

---

3. Run on YARN with spark-submit

Follow the steps below if you need to run BigDL with spark-submit.

• Pack the current active conda environment to environment.tar.gz (you can use any name you like) in the current working directory:

  conda pack -o environment.tar.gz
  

• You need to write your BigDL program as a Python script. In the script, you need to call init_orca_context at the very beginning of your code and specify "cluster_mode" to be spark-submit:

  from bigdl.orca import init_orca_context
  
  sc = init_orca_context(cluster_mode="spark-submit")
  

• Use bigdl-submit to submit your BigDL program (e.g. script.py). You can adjust the configurations according to your cluster settings. Note that if environment.tar.gz is not under the same directory with script.py, you may need to modify its path in --archives in the running command below.

  For yarn-cluster mode:

  bigdl-submit \
      --conf spark.yarn.appMasterEnv.PYSPARK_PYTHON=environment/bin/python \
      --conf spark.executorEnv.PYSPARK_PYTHON=environment/bin/python \
      --master yarn \
      --deploy-mode cluster \
      --executor-memory 10g \
      --driver-memory 10g \
      --executor-cores 8 \
      --num-executors 2 \
      --archives environment.tar.gz#environment \
      script.py
  

  Note: For yarn-cluster, the Spark driver is running in a YARN container as well and thus both the driver and executors will use the Python interpreter in environment.tar.gz. If you want to operate HDFS as some certain user, you can add spark.yarn.appMasterEnv.HADOOP_USER_NAME=username to SparkConf.

  For yarn-client mode:

  PYSPARK_PYTHON=environment/bin/python bigdl-submit \
      --master yarn \
      --deploy-mode client \
      --executor-memory 10g \
      --driver-memory 10g \
      --executor-cores 8 \
      --num-executors 2 \
      --archives environment.tar.gz#environment \
      script.py
  

  Note: For yarn-client, the Spark driver is running on local and it will use the Python interpreter in the current active conda environment while the executors will use the Python interpreter in environment.tar.gz.