open-nomad/website/source/guides/spark/resource.html.md
2017-06-28 17:10:05 -07:00

3.1 KiB

layout page_title sidebar_current description
guides Apache Spark Integration - Resource Allocation guides-spark-resource Learn how to configure resource allocation for your Spark applications.

Resource Allocation

Resource allocation can be configured using a job template or through configuration properties. Here is a sample template in HCL syntax (this would need to be converted to JSON):

job "template" {
  group "group-name" {

    task "executor" {
      meta {
        "spark.nomad.role" = "executor"
      }

      resources {
        cpu = 2000
        memory = 2048
        network {
          mbits = 100
        }
      }
    }
  }
}

Resource-related configuration properties are covered below.

Memory

The standard Spark memory properties will be propagated to Nomad to control task resource allocation: spark.driver.memory (set by --driver-memory) and spark.executor.memory (set by --executor-memory). You can additionally specify spark.nomad.shuffle.memory to control how much memory Nomad allocates to shuffle service tasks.

CPU

Spark sizes its thread pools and allocates tasks based on the number of CPU cores available. Nomad manages CPU allocation in terms of processing speed rather than number of cores. When running Spark on Nomad, you can control how much CPU share Nomad will allocate to tasks using the spark.nomad.driver.cpu (set by --driver-cpu), spark.nomad.executor.cpu (set by --executor-cpu) and spark.nomad.shuffle.cpu properties. When running on Nomad, executors will be configured to use one core by default, meaning they will only pull a single 1-core task at a time. You can set the spark.executor.cores property (set by --executor-cores) to allow more tasks to be executed concurrently on a single executor.

Network

Nomad does not restrict the network bandwidth of running tasks, bit it does allocate a non-zero number of Mbit/s to each task and uses this when bin packing task groups onto Nomad clients. Spark defaults to requesting the minimum of 1 Mbit/s per task, but you can change this with the spark.nomad.driver.networkMBits, spark.nomad.executor.networkMBits, and spark.nomad.shuffle.networkMBits properties.

Log rotation

Nomad performs log rotation on the stdout and stderr of its tasks. You can configure the number number and size of log files it will keep for driver and executor task groups using spark.nomad.driver.logMaxFiles and spark.nomad.executor.logMaxFiles.

Next Steps

Learn how to dynamically allocate Spark executors.