Step 2: Configure Drill-on-YARN

To configure Drill to run as a YARN application, modify the $DRILL_SITE/drill-on-yarn.conf cluster configuration file to suit the needs of your cluster. This file is a “starter” configuration file that corresponds to the simplest Drill cluster. The drill-on-yarn.conf file is in the same HOCON format as drill-override.conf.

Consult the $DRILL_HOME/conf/drill-on-yarn-example.conf file as an example. However, do not just copy the example file. Instead, copy only the specific configuration settings that you need; the others will automatically take the Drill-defined default values.

NOTE Make sure that resources can accommodate the Drill memory, CPU, and disk requirements.

The following sections list the configuration settings required to run Drill under YARN:

Drill Memory Settings

The following configuration sets the Java heap size and amount of direct memory the node can allocate to Drill:

drillbit: { heap: "<value>" max-direct-memory: "<value>" }

When you add the configuration, use the same values set in the following parameters of the drill-env.sh file, if you did not remove these lines when you modified drill-env.sh:

export DRILL_MAX_DIRECT_MEMORY=${DRILL_MAX_DIRECT_MEMORY:-"<value>"}
export DRILL_HEAP=${DRILL_HEAP:-"<value>"}

Drill-on-YARN copies these values into the environment variables when launching each Drillbit. Drill also uses additional JVM memory. For example, Drill uses a code cache to hold classes generated at runtime. The default size of the cache is 1 GB:

drillbit: { code-cache: "1G" }

Typically, you do not need to change the code cache size, but you must account for it when computing the YARN container size.

YARN Container Size

The following configuration sets the YARN container size required to run Drill as a YARN application.

drillbit: {
   memory-mb: 14336
 }

The default value is 14GB. Typically, this size is the sum of the heap and direct memory. However, if you use custom libraries that perform their own memory allocation, or launch sub-processes, you must account for that memory usage as well. Note that YARN memory is expressed in MB. To compute the container size, start with the values used for the heap, direct memory, and code cache settings, as shown in the following example:

drillbit: {
    heap: "4G"
    max-direct-memory: "8G"
    code-cache: "1G"
    memory-mb: 14336
  }

The values shown above are the Drill defaults. You may use larger values. Although the three values account for the bulk of Drill memory, the JVM itself also has a certain overhead. Assume that the overhead is about 1 GB, though the amount varies depending on the workload.

Add the four values together to get a memory requirement in GB.

Total memory = 8G + 4G + 1G + 1G = 14G

YARN sizes containers in megabytes. Convert GB to MB:

Container size = 14G * 1024 = 14336 MB

Set this size in drill-on-yarn.conf:

drillbit: { memory-mb: 14336 }

CPU

The following configuration sets the CPU to allocate to Drill:

drillbit: { vcores: <value> }

Drill is a CPU-intensive operation and greatly benefits from each additional core. YARN does not limit the number of cores used by an application. Rather, this number reports to YARN the average CPU usage of Drill so that YARN can use the number when deciding how many other applications to run on the same node.

Drillbit Cluster Configuration

The following configuration sets the cluster group:

cluster: [ { name: "drillbits" type: "basic" count: 1 } ]

Drill-on-YARN uses the concept of a “cluster group” of Drillbits to describe the set of drillbits to launch. Currently, only the “basic” type of group is supported. A basic group launches drillbits anywhere in the YARN cluster where a container is available. For a basic group, specify the group type and the number of drillbits to launch.

NOTE The syntax says that cluster is a list that contains cluster group objects contained in braces. Drill currently supports just one group.

The name is optional. It appears in the Application Master web UI. Type must be set to “basic.” Set the count to the number of hosts on which Drill is to run at launch time. You can resize the Drill cluster after the cluster is launched.

YARN Queue Labels

The following configuration sets the YARN queue labels that identify the cluster nodes that run Drill:

yarn: { queue: "<queue_name>" }

The MapR distribution of YARN provides queue labels for assigning YARN applications to specific queues. See Label-Based Scheduling for YARN Applications. You can use queue labels with Drill to identify the YARN queue that should run Drill.

To use queue labels, complete the following steps:

Create a node label.
Assign the label to the nodes that are to run Drill.
Create a Drill-specific queue that uses the node label.
Configure Drill-on-YARN to use the queue.

Suppose you create a queue called “drill.” Setting the following configuration causes Drill-on-YARN to launch through the drill queue:

yarn: { queue: "drill" }

Set queue to the name a name of your choice. When Drill-on-YARN launches, both the Application Master and drillbits run only on nodes with the same node label as the queue.

DFS Location

The following configuration sets the dfs location:

dfs: { app-dir: "/user/drill" }

Drill copies the archive in to the MapR filesystem in a location you provide. The default is /user/drill, however you can specify a different location. You do not have to specify the MapR File System connection information; MapR defines this automatically.