This is a getting started guide to XGBoost4J-Spark on Apache Hadoop YARN supporting GPU scheduling. At the end of this guide, the reader will be able to run a sample Apache Spark application that runs on NVIDIA GPUs.
- Apache Spark 3.0.1+ running on YARN supporting GPU scheduling. (e.g.: Spark 3.0.1, Hadoop-Yarn 3.1.0)
- Hardware Requirements
- NVIDIA Pascal™ GPU architecture or better
- Multi-node clusters with homogenous GPU configuration
- Software Requirements
- Ubuntu 18.04, 20.04/CentOS7, CentOS8
- CUDA 11.0-11.4
- NVIDIA driver compatible with your CUDA
- NCCL 2.7.8
The number of GPUs per NodeManager dictates the number of Spark executors that can run in that NodeManager. Additionally, cores per Spark executor and cores per Spark task must match, such that each executor can run 1 task at any given time.
For example: if each NodeManager has 4 GPUs, there should be 4 or less executors running on each NodeManager, and each executor should run 1 task (e.g.: A total of 4 tasks running on 4 GPUs). In order to achieve this, you may need to adjust spark.task.cpus and spark.executor.cores to match (both set to 1 by default). Additionally, we recommend adjusting executor-memory to divide host memory evenly amongst the number of GPUs in each NodeManager, such that Spark will schedule as many executors as there are GPUs in each NodeManager.
We use SPARK_HOME to point to the cluster's Apache Spark installation. And as to how to enable GPU scheduling and isolation for Yarn, please refer to here.
Make sure you have prepared the necessary packages and dataset by following this guide
Create a directory in HDFS, and copy:
[xgboost4j_spark]$ hadoop fs -mkdir /tmp/xgboost4j_spark
[xgboost4j_spark]$ hadoop fs -copyFromLocal ${SPARK_XGBOOST_DIR}/mortgage/* /tmp/xgboost4j_sparkVariables required to run spark-submit command:
# location where data was downloaded
export DATA_PATH=hdfs:/tmp/xgboost4j_spark/data
# spark deploy mode (see Apache Spark documentation for more information)
export SPARK_DEPLOY_MODE=cluster
# run a single executor for this example to limit the number of spark tasks and
# partitions to 1 as currently this number must match the number of input files
export SPARK_NUM_EXECUTORS=1
# spark driver memory
export SPARK_DRIVER_MEMORY=4g
# spark executor memory
export SPARK_EXECUTOR_MEMORY=8g
# example class to use
export EXAMPLE_CLASS=com.nvidia.spark.examples.mortgage.GPUMain
# tree construction algorithm
export TREE_METHOD=gpu_histRun spark-submit:
${SPARK_HOME}/bin/spark-submit \
--conf spark.plugins=com.nvidia.spark.SQLPlugin \
--conf spark.rapids.memory.gpu.pooling.enabled=false \
--conf spark.executor.resource.gpu.amount=1 \
--conf spark.task.resource.gpu.amount=1 \
--conf spark.executor.resource.gpu.discoveryScript=./getGpusResources.sh \
--files $SPARK_HOME/examples/src/main/scripts/getGpusResources.sh \
--jars ${CUDF_JAR},${RAPIDS_JAR} \
--master yarn \
--deploy-mode ${SPARK_DEPLOY_MODE} \
--num-executors ${SPARK_NUM_EXECUTORS} \
--driver-memory ${SPARK_DRIVER_MEMORY} \
--executor-memory ${SPARK_EXECUTOR_MEMORY} \
--class ${EXAMPLE_CLASS} \
${SAMPLE_JAR} \
-dataPath=train::${DATA_PATH}/mortgage/csv/train/mortgage_train_merged.csv \
-dataPath=trans::${DATA_PATH}/mortgage/csv/test/mortgage_eval_merged.csv \
-format=csv \
-numWorkers=${SPARK_NUM_EXECUTORS} \
-treeMethod=${TREE_METHOD} \
-numRound=100 \
-maxDepth=8 In the stdout driver log, you should see timings* (in seconds), and the accuracy metric:
--------------
==> Benchmark: Elapsed time for [Mortgage GPU train csv stub Unknown Unknown Unknown]: 29.642s
--------------
--------------
==> Benchmark: Elapsed time for [Mortgage GPU transform csv stub Unknown Unknown Unknown]: 21.272s
--------------
--------------
==> Benchmark: Accuracy for [Mortgage GPU Accuracy csv stub Unknown Unknown Unknown]: 0.9874184013493451
--------------
If you are running this example after running the GPU example above, please set these variables, to set both training and testing to run on the CPU exclusively:
# location where data was downloaded
export DATA_PATH=hdfs:/tmp/xgboost4j_spark/data
# spark deploy mode (see Apache Spark documentation for more information)
export SPARK_DEPLOY_MODE=cluster
# run a single executor for this example to limit the number of spark tasks and
# partitions to 1 as currently this number must match the number of input files
export SPARK_NUM_EXECUTORS=1
# spark driver memory
export SPARK_DRIVER_MEMORY=4g
# spark executor memory
export SPARK_EXECUTOR_MEMORY=8g
# example class to use
export EXAMPLE_CLASS=com.nvidia.spark.examples.mortgage.CPUMain
# tree construction algorithm
export TREE_METHOD=histThis is the same command as for the GPU example, repeated for convenience:
${SPARK_HOME}/bin/spark-submit \
--master yarn \
--deploy-mode ${SPARK_DEPLOY_MODE} \
--num-executors ${SPARK_NUM_EXECUTORS} \
--driver-memory ${SPARK_DRIVER_MEMORY} \
--executor-memory ${SPARK_EXECUTOR_MEMORY} \
--class ${EXAMPLE_CLASS} \
${SAMPLE_JAR} \
-dataPath=train::${DATA_PATH}/mortgage/csv/train/mortgage_train_merged.csv \
-dataPath=trans::${DATA_PATH}/mortgage/csv/test/mortgage_eval_merged.csv \
-format=csv \
-numWorkers=${SPARK_NUM_EXECUTORS} \
-treeMethod=${TREE_METHOD} \
-numRound=100 \
-maxDepth=8 In the stdout driver log, you should see timings* (in seconds), and the accuracy metric:
--------------
==> Benchmark: Elapsed time for [Mortgage CPU train csv stub Unknown Unknown Unknown]: 286.398s
--------------
--------------
==> Benchmark: Elapsed time for [Mortgage CPU transform csv stub Unknown Unknown Unknown]: 49.836s
--------------
--------------
==> Benchmark: Accuracy for [Mortgage CPU Accuracy csv stub Unknown Unknown Unknown]: 0.9873709530950067
--------------
* The timings in this Getting Started guide are only illustrative. Please see our release announcement for official benchmarks.