本文属于机器翻译版本。若本译文内容与英语原文存在差异，则一律以英文原文为准。

# 将适用于 Apache Spark 的 RAPIDS Accelerator 与 Amazon EMR on EKS 结合使用
<a name="tutorial-spark-rapids"></a>

使用 Amazon EMR on EKS，您可以为适用于 Apache Spark 的 Nvidia RAPIDS Accelerator 运行任务。本教程介绍如何在 EC2 图形处理单元（GPU）实例类型上使用 RAPIDS 运行 Spark 任务。本教程使用以下版本：
+ Amazon EMR on EKS 发行版 6.9.0 及更高版本
+ Apache Spark 3.x

您可以借助[适用于 Apache Spark 的 Nvidia RAPIDS Accelerator](https://docs.nvidia.com/spark-rapids/user-guide/latest/overview.html) 插件，使用 Amazon EC2 GPU 实例类型加速 Spark。当您结合使用这些技术时，您可以加速数据科学管道，而无需更改任何代码。这可以减少数据处理和模型训练所需的运行时间。通过在更短的时间内完成更多工作，降低基础设施成本。

在开始之前，请确保您具有以下资源。
+ Amazon EMR on EKS 虚拟集群
+ 包含支持 GPU 的节点组的 Amazon EKS 集群

Amazon EKS 虚拟集群是 Amazon EKS 集群上 Kubernetes 命名空间的注册句柄，由 Amazon EMR on EKS 管理。该句柄允许 Amazon EMR 使用 Kubernetes 命名空间作为运行任务的目标。有关如何设置虚拟集群的更多信息，请参阅本指南中的 [设置 Amazon EMR on EKS](setting-up.md)。

您必须使用具有 GPU 实例的节点组配置 Amazon EKS 虚拟集群。您必须使用 Nvidia 设备插件配置节点。请参阅 [managed node groups](https://docs.amazonaws.cn/eks/latest/userguide/managed-node-groups.html)（托管节点组）以了解更多信息。

要将 Amazon EKS 集群配置为添加支持 GPU 的节点组，请执行以下步骤：

**添加支持 GPU 的节点组**

1. 使用以下 [create-nodegroup](https://docs.amazonaws.cn/cli/latest/reference/eks/create-nodegroup.html) 命令创建支持 GPU 的节点组。请务必为 Amazon EKS 集群替换正确的参数。使用支持 Spark RAPIDS 的实例类型，例如 P4、P3、G5 或 G4dn。

   ```
   aws eks create-nodegroup \
    --cluster-name EKS_CLUSTER_NAME \
    --nodegroup-name NODEGROUP_NAME \
    --scaling-config minSize=0,maxSize=5,desiredSize=2 CHOOSE_APPROPRIATELY \
    --ami-type AL2_x86_64_GPU \
    --node-role NODE_ROLE \
    --subnets SUBNETS_SPACE_DELIMITED  \
    --remote-access ec2SshKey= SSH_KEY \
    --instance-types GPU_INSTANCE_TYPE \
    --disk-size DISK_SIZE \
    --region AWS_REGION
   ```

1. 在集群中安装 Nvidia 设备插件，以在集群的每个节点 GPUs 上发出数量，并在集群中运行支持 GPU 的容器。运行以下命令以安装插件：

   ```
   kubectl apply -f https://raw.githubusercontent.com/NVIDIA/k8s-device-plugin/v0.9.0/nvidia-device-plugin.yml
   ```

1. 要验证集群的每个节点上 GPUs 有多少可用节点，请运行以下命令：

   ```
   kubectl get nodes  "-o=custom-columns=NAME:.metadata.name,GPU:.status.allocatable.nvidia\.com/gpu"
   ```

**运行 Spark RAPIDS 任务**

1. 将 Spark RAPIDS 任务提交到 Amazon EMR on EKS 集群。以下代码显示了用于启动任务的命令示例。首次运行任务时，可能需要几分钟下载映像并将其缓存到节点上。

   ```
   aws emr-containers start-job-run \
   --virtual-cluster-id VIRTUAL_CLUSTER_ID \
   --execution-role-arn JOB_EXECUTION_ROLE \
   --release-label emr-6.9.0-spark-rapids-latest \
   --job-driver '{"sparkSubmitJobDriver": {"entryPoint": "local:///usr/lib/spark/examples/jars/spark-examples.jar","entryPointArguments":  ["10000"], "sparkSubmitParameters":"--class org.apache.spark.examples.SparkPi "}}' \
   ---configuration-overrides '{"applicationConfiguration": [{"classification": "spark-defaults","properties": {"spark.executor.instances": "2","spark.executor.memory": "2G"}}],"monitoringConfiguration": {"cloudWatchMonitoringConfiguration": {"logGroupName": "LOG_GROUP _NAME"},"s3MonitoringConfiguration": {"logUri": "LOG_GROUP_STREAM"}}}'
   ```

1. 要验证 Spark RAPIDS Accelerator 是否已启用，请检查 Spark 驱动程序日志。这些日志存储在您运行`start-job-run`命令时指定的 S3 位置 CloudWatch 或位置。以下示例大致显示了日志行的具体形式：

   ```
   22/11/15 00:12:44 INFO RapidsPluginUtils: RAPIDS Accelerator build: {version=22.08.0-amzn-0, user=release, url=, date=2022-11-03T03:32:45Z, revision=, cudf_version=22.08.0, branch=}
   22/11/15 00:12:44 INFO RapidsPluginUtils: RAPIDS Accelerator JNI build: {version=22.08.0, user=, url=https://github.com/NVIDIA/spark-rapids-jni.git, date=2022-08-18T04:14:34Z, revision=a1b23cd_sample, branch=HEAD}
   22/11/15 00:12:44 INFO RapidsPluginUtils: cudf build: {version=22.08.0, user=, url=https://github.com/rapidsai/cudf.git, date=2022-08-18T04:14:34Z, revision=a1b23ce_sample, branch=HEAD}
   22/11/15 00:12:44 WARN RapidsPluginUtils: RAPIDS Accelerator 22.08.0-amzn-0 using cudf 22.08.0.
   22/11/15 00:12:44 WARN RapidsPluginUtils: spark.rapids.sql.multiThreadedRead.numThreads is set to 20.
   22/11/15 00:12:44 WARN RapidsPluginUtils: RAPIDS Accelerator is enabled, to disable GPU support set `spark.rapids.sql.enabled` to false.
   22/11/15 00:12:44 WARN RapidsPluginUtils: spark.rapids.sql.explain is set to `NOT_ON_GPU`. Set it to 'NONE' to suppress the diagnostics logging about the query placement on the GPU.
   ```

1. 要查看将在 GPU 上运行的操作，请执行以下步骤以启用额外的日志记录。请注意“`spark.rapids.sql.explain : ALL`”配置。

   ```
   aws emr-containers start-job-run \
   --virtual-cluster-id VIRTUAL_CLUSTER_ID \
   --execution-role-arn JOB_EXECUTION_ROLE \
   --release-label emr-6.9.0-spark-rapids-latest \
   --job-driver '{"sparkSubmitJobDriver": {"entryPoint": "local:///usr/lib/spark/examples/jars/spark-examples.jar","entryPointArguments":  ["10000"], "sparkSubmitParameters":"--class org.apache.spark.examples.SparkPi "}}' \
   ---configuration-overrides '{"applicationConfiguration": [{"classification": "spark-defaults","properties": {"spark.rapids.sql.explain":"ALL","spark.executor.instances": "2","spark.executor.memory": "2G"}}],"monitoringConfiguration": {"cloudWatchMonitoringConfiguration": {"logGroupName": "LOG_GROUP_NAME"},"s3MonitoringConfiguration": {"logUri": "LOG_GROUP_STREAM"}}}'
   ```

   上一个命令是使用 GPU 的任务示例。其输出类似于以下示例。请参阅此密钥以帮助了解输出：
   + `*`：标记适用于 GPU 的操作
   + `!`：标记无法在 GPU 上运行的操作
   + `@`：标记适用于 GPU 但无法运行的操作，因为其所处的计划无法在 GPU 上运行

   ```
    22/11/15 01:22:58 INFO GpuOverrides: Plan conversion to the GPU took 118.64 ms
    22/11/15 01:22:58 INFO GpuOverrides: Plan conversion to the GPU took 4.20 ms
    22/11/15 01:22:58 INFO GpuOverrides: GPU plan transition optimization took 8.37 ms
    22/11/15 01:22:59 WARN GpuOverrides:
       *Exec <ProjectExec> will run on GPU
         *Expression <Alias> substring(cast(date#149 as string), 0, 7) AS month#310 will run on GPU
           *Expression <Substring> substring(cast(date#149 as string), 0, 7) will run on GPU
             *Expression <Cast> cast(date#149 as string) will run on GPU
         *Exec <SortExec> will run on GPU
           *Expression <SortOrder> date#149 ASC NULLS FIRST will run on GPU
           *Exec <ShuffleExchangeExec> will run on GPU
             *Partitioning <RangePartitioning> will run on GPU
               *Expression <SortOrder> date#149 ASC NULLS FIRST will run on GPU
             *Exec <UnionExec> will run on GPU
               !Exec <ProjectExec> cannot run on GPU because not all expressions can be replaced
                 @Expression <AttributeReference> customerID#0 could run on GPU
                 @Expression <Alias> Charge AS kind#126 could run on GPU
                   @Expression <Literal> Charge could run on GPU
                 @Expression <AttributeReference> value#129 could run on GPU
                 @Expression <Alias> add_months(2022-11-15, cast(-(cast(_we0#142 as bigint) + last_month#128L) as int)) AS date#149 could run on GPU
                   ! <AddMonths> add_months(2022-11-15, cast(-(cast(_we0#142 as bigint) + last_month#128L) as int)) cannot run on GPU because GPU does not currently support the operator class org.apache.spark.sql.catalyst.expressions.AddMonths
                     @Expression <Literal> 2022-11-15 could run on GPU
                     @Expression <Cast> cast(-(cast(_we0#142 as bigint) + last_month#128L) as int) could run on GPU
                       @Expression <UnaryMinus> -(cast(_we0#142 as bigint) + last_month#128L) could run on GPU
                         @Expression <Add> (cast(_we0#142 as bigint) + last_month#128L) could run on GPU
                           @Expression <Cast> cast(_we0#142 as bigint) could run on GPU
                             @Expression <AttributeReference> _we0#142 could run on GPU
                           @Expression <AttributeReference> last_month#128L could run on GPU
   ```