Running applications using CDS 3.2.3 with GPU SupportCDS 3.2.3 with GPU Support

Log on to the node where you want to run the job.
Run the following command to launch spark3-shell:
```
spark3-shell  --conf "spark.rapids.sql.enabled=true" \
              --conf "spark.executor.memoryOverhead=5g"
```
where
--conf spark.rapids.sql.enabled=true
enables the following environment variables for GPUs:
```
"spark.task.resource.gpu.amount" - sets GPU resource amount per task

"spark.rapids.sql.concurrentGpuTasks" - sets the number of concurrent tasks per GPU
"spark.sql.files.maxPartitionBytes" - sets the input partition size for DataSource API, The recommended value is "256m".
"spark.locality.wait" - controls how long Spark waits to obtain better locality for tasks. 
"spark.sql.adaptive.enabled" - enables Adaptive Query Execution.
"spark.rapids.memory.pinnedPool.size" - sets the amount of pinned memory allocated per host.
"spark.sql.adaptive.advisoryPartitionSizeInBytes" - sets the advisory size in bytes of the shuffle partition during adaptive optimization.
```
For example,
```
$SPARK_HOME/bin/spark3-shell \
       --conf spark.task.resource.gpu.amount=2 \
       --conf spark.rapids.sql.concurrentGpuTasks=2 \
       --conf spark.sql.files.maxPartitionBytes=256m \
       --conf spark.locality.wait=0s \
       --conf spark.sql.adaptive.enabled=true \
       --conf spark.rapids.memory.pinnedPool.size=2G \
       --conf spark.sql.adaptive.advisoryPartitionSizeInBytes=1g
```
--conf "spark.executor.memoryOverhead=5g"

sets the amount of additional memory to be allocated per executor process
note
cuDF uses a Just-In-Time (JIT) compilation approach for some kernels, and the JIT process can add a few seconds to query wall-clock time. You are recommended to set a JIT cache path to speed up subsequent invocations with: --conf spark.executorEnv.LIBCUDF_KERNEL_CACHE_PATH=[local path]. The path should be local to the executor (not HDFS) and not shared between different cluster users in a multi-tenant environment. For example, the path may be in /tmp: (/tmp/cudf-[***USER***]). If the/tmp directory is not writable, consult your system administrator to find a path that is.
You can override these configuration settings both from the command line and from code. For more information on environment variables, see the NVIDIA spark-rapids documentation and the Spark SQL Performance Tuning Guide.

Run a job in spark3-shell.

For example:

scala> val df = sc.makeRDD(1 to 100000000, 6).toDF
df: org.apache.spark.sql.DataFrame = [value: int]

scala>val df2 = sc.makeRDD(1 to 100000000, 6).toDF
df2: org.apache.spark.sql.DataFrame = [value: int]

scala> df.select($"value" as "a").join(df2select($"value" as "b"), $"a" === $"b").count
res0: Long = 100000000

You can verify that the job run used GPUs, by logging on to the Yarn UI v2 to review the execution plan and the performance of your spark3-shell application:
Select the Applications tab then select your [spark3-shell application]. Select ApplicationMaster > SQL > count at <console>:28 to see the execution plan.

Running a Spark job using CDS 3.2.3 with GPU Support with UCX enabled

Log on to the node where you want to run the job.
Run the following command to launch spark3-shell:
```
spark3-shell  --conf "spark.rapids.sql.enabled=true" \
              --conf "spark.executor.memoryOverhead=5g"
              --rapids-shuffle=true
```
where
--rapids-shuffle=true
makes the following configuration changes for UCX:
```
spark.shuffle.manager=com.nvidia.spark.rapids.spark320cdh.RapidsShuffleManager
spark.executor.extraClassPath=/opt/cloudera/parcels/SPARK3/lib/spark3/rapids-plugin/*
spark.executorEnv.UCX_ERROR_SIGNALS=
spark.executorEnv.UCX_MEMTYPE_CACHE=n
```
important
If you using Cloudera Runtime 7.1.7 Service Pack 1 or 2, you must provide the value of spark.shuffle.manager as com.nvidia.spark.rapids.spark321cdh.RapidsShuffleManager.
For more information on environment variables, see the NVIDIA spark-rapids documentation.
Run a job in spark3-shell.