Enabling and configuring the Spark shuffle service

When your instances (instance groups, Anaconda distribution instances, and application instances) are deployed to a local file system , the shuffle service is enabled by default. When your instances are is deployed to a shared file system, the shuffle service is disabled by default to increase application performance.

1. In the Basic Settings, select the Spark version that the instance group must use.
2. Click Configuration to define the following shuffle service settings:

   Spark parameter	Description
   spark.shuffle.service.port	Define an exclusive port for use by the Spark shuffle service (default 7337). To ensure a unique environment for each instance group, the default port number increments by 1 for each instance group that you subsequently create. So the shuffle service for your second instance group runs by default on port 7338, the next on port 7339, and so on. However, when you create a instance group by using RESTful APIs, the shuffle service uses port 7337 by default for all instance groups. Ensure that the shuffle service port is not already used by the system or by processes external to IBM® Spectrum Conductor. This port is not used on all hosts that start the shuffle service; you need to only check hosts in the shuffle service resource group.
   spark.local.dir	When your instances (instance groups, Anaconda distribution instances, and application instances) are deployed to a local file system (with the shuffle service enabled), specify a directory for scratch space (including map output files and RDDs) that is on a fast, local disk in your system (default is /tmp). It can also be a comma-separated list of multiple directories on different disks. If spark.local.dir is specified to an existing directory, ensure that the execution user for the instance group and the execution user for the Spark shuffle service consumer have read/write/execute (rwx) permissions for that directory. If you change the default consumer for each component, the consumer execution user of the driver and executor processes for any submitted applications, the Spark batch master service, and the Spark notebook master service must have execute (x) permission to the spark.local.dir directory. Note: With the shuffle service enabled, you must manually clean up data under the spark.local.dir location when the instance group is removed.
   SPARK_EGO_LOGSERVICE_PORT	Specify the UI port for the EGO log service (default 28082). To ensure a unique environment for each instance group, the default port number increments by 1 for each instance group that you subsequently create. So the log service for your second instance group runs by default on port 28083, the next on port 28084, and so on. However, when you create a instance group by using RESTful APIs, the log service uses port 28082 by default for all instance groups. Ensure that the log service UI port is not already used by the system or by processes external to IBM Spectrum Conductor. This port is not used on all hosts that start the log service; you need to only check hosts in the shuffle service resource group.

3. If you want to change the default shuffle service consumer, click Edit detailed consumer assignments, then the consumer for the shuffle service.
   • To select an existing child consumer, click Select an existing child consumer, select the consumer under the top-level consumer, and click Select.
   • To create a new consumer, click Create a new consumer under top_level_consumer, enter a consumer name, and click Create.

   If you plan to use exclusive slot allocation, the shuffle service consumer and the Spark executors consumer must be different.
4. If you want the shuffle service and executors to use different resource groups, in the Resource Groups and Plans section, select a different resource group for the shuffle service. Both resource groups must use the same set of hosts; otherwise, applications that are submitted to this instance group fail.