This section provides an overview of configuration recommendations to be applied to the Hadoop cluster from the 

YARN manages cluster resources (CPU and memory) by running all processes within allocated containers. Containers restrict the resources available to its process(es). Processes are monitored and killed if they overrun the container allocation.

• Multiple containers can run on a cluster node (if available resources permit).
• A job can request and use multiple containers across the cluster.
• Container requests specify virtual CPU (cores) and memory (in MB).

YARN configuration specifies:

• Per Cluster Node: Available virtual CPUs and memory per cluster node
• Per Container: virtual CPUs and memory for each container

The following parameters are available in yarn-site.xml:

Spark Tuning Overview

Spark processes run multiple executors per job. Each executor must run within a YARN container. Therefore, resource requests must fit within YARN’s container limits.

Like YARN containers, multiple executors can run on a single node. More executors provide additional computational power and decreased runtime.

Dynamic allocation

Spark’s dynamic allocation adjusts the number of executors to launch based on the following:

• job size

• job complexity

• available resources

For more information, see https://spark.apache.org/docs/latest/configuration.html#dynamic-allocation.

Per-executor allocations

The per-executor resource request sizes can be specified by setting the following properties in the spark.props section :

A single special process, the application driver, also runs in a container. Its resources are specified in the spark.props section:

Spark Performance Considerations

Optimizing "Small" Joins

Broadcast, or map-side, joins materialize one side of the join and send it to all executors to be stored in memory. This technique can significantly accelerate joins by skipping the sort and shuffle phases during a "reduce" operation. However, there is also a cost in communicating the table to all executors. Therefore, only "small" tables should be considered for broadcast join. The definition of "small" is set by the spark.sql.autoBroadcastJoinThreshold parameter which can be added to the spark.props section of 

Checkpointing

In Spark's driver process, the transformation pipeline is compiled down to Spark code and optimized. This process can sometimes fail or take and an inordinately long time. By checkpointing the execution, Spark is forced to materialize the current table (in memory or on disk), thereby simplifying the segments that are optimized. While checkpointing can incur extra cost due to this materialization, it can also reduce end-to-end execution time by speeding up the compilation and optimization phases and by reusing materialized columns downstream.

Limiting Resource Utilization of Spark Jobs

With Spark's dynamic allocation, each job's resource utilization can be limited by setting the maximum number of executors per job. Set spark.dynamicAllocation.maxExecutors in the spark.props section of 

spark.dynamicAllocation.maxExecutors * (spark.driver.memory + spark.executor.memory)

The maximum number of cores used per job is given (exactly) by:

spark.dynamicAllocation.maxExecutors * (spark.driver.cores + spark.executor.cores)

To limit the overall cluster utilization of 

Tuning Recommendations

The following configuration settings can be applied through 

The specified configuration allows, maximally, the following Spark configuration per node:

Spark Job Property Overrides

You can enable a set of Spark properties that users are permitted to override on individual jobs. For more information, see Enable Spark Job Overrides.