SPARK running explained - 3

SPARK running explained - 2

SPARK running explained - 1

1.       YARN Cluster Manager – Basic YARN architecture is described as below, and it is similar to Spark Standalone Cluster Manager. Main components –

a.       Resource manager – (Just like Spark Master process)

b.      Node manager – (similar to Spark’s worker processes)

Unlike running on Spark’s standalone cluster, applications on YARN run in containers (JVM processes to which CPU and memory resources are granted).

There is an “Application Master” for each application, running in its own container, it’s responsible for requesting application resources from Resource Manager.

Node managers track resources used by containers and report to the resource manager.

Below depicts the Spark application (cluster-deploy mode) running on YARN cluster with 2 nodes-

1.       Client submit application to Resource Manager

2.       Resource Manger asks one node manager to allocate  container for Application Master

3.       The node manager launches a container for Application Master, Spark Driver runs in Application master.

4.       Application Master asks Resource Manager for more containers for executors.

5.       After Resource manager grants resources then Application Master asks node managers to launch new containers for executors.

6.       Node managers starts executors on behalf of Spark Application master.

7.       After that, executors and driver communicate independently.

Unlike Spark’s workers, YARN’s node managers can launch more than one container (executor) per application.

                3 possible modes of YARN are described as below –

1.       Standalone (local) mode - Runs as a single Java process

2.       Pseudo-distributed mode - Runs all Hadoop daemons (several Java processes) on a single machine

3.       Fully distributed mode – Runs various Java processed on multiple machines

Resource scheduling in YARN- There are 3 main schedulers that can be plugged in –

1.       FIFO scheduler - If two applications require the same resources, the first application that requests them will be first served (FIFO).

2.       CAPACITY scheduler – It guarantees capacity for different organization using same cluster. The main unit of resources scheduled by YARN is a queue. Each queue’s capacity determines the percentage of cluster resources that can be used by applications submitted to it. A hierarchy of queues can be set up to reflect a hierarchy of capacity requirements by organizations, so that sub-queues (sub-organizations) can share the resources of a single queue and thus not affect others. In a single queue, the resources are scheduled in FIFO fashion. If enabled, capacity scheduling can be elastic, meaning it allows organizations to use any excess capacity not used by others. But, preemption isn’t supported.

3.       FAIR Scheduler - The fair scheduler tries to assign resources in such a way that all applications get (on average) an equal share. Like the capacity scheduler, it also organizes applications into queues. This scheduler supports application priorities and minimum capacity requirements. It enables preemption, meaning when an application demands resources, the fair scheduler can take some resources from other running applications.

Configuring resources for Spark jobs

Set following properties –

1.       --num-executors—Changes the number of executors

2.       --executor-cores—Changes the number of cores per executor

Driver Memory can be set –

1.       --driver-memory command-line

2.       “spark.driver.memory” configuration parameter

3.       SPARK_DRIVER_MEMORY environment variable

Spark Executor memory can be set-

1.       “spark.executor.memory” configuration parameter

2.       SPARK_EXECUTOR_MEMORY environment variable

3.       --executor-memory command-line

4.       “spark.executor.memoryOverhead” - Additional parameter, determines additional memory beyond the Java Heap that will be available to YARN containers running Spark executors. This memory is for JVM process itself

Note - If your executor uses more memory than spark.executor.memory + spark.executor.memoryOverhead YARN will shut down the container, and your jobs will repeatedly fail.

Failing to set spark.executor.memoryOverhead to a sufficiently high value can lead to problems that are hard to diagnose. Make sure to specify at least 1024 MB.

Thus, overall container memory can be segmented in to –

1.       Memory Overhead

2.       Spark Memory

a.       Storage memory (including safety fraction)

b.      Shuffle memory (including safety fraction)

c.       Rest of Heap for Java objects

In cluster-deploy mode – “spark.yarn.driver.memoryOverhead” – determines the memory overhead of the driver’s container

In client-deploy mode -“spark.yarn.am.memoryOverhead” - determines the memory overhead of the application master

Other YARN configuration parameters –

1.       “yarn.scheduler.maximum-allocation-mb” = Determines the upper memory limit of YARN containers. The resource manager won’t allow allocation of larger amounts of memory. The default value is 8192 MB

2.       “yarn.scheduler.minimum-allocation-mb” = Determines the minimum amount of memory the resource manager can allocate. The resource manager allocates memory only in multiples of this parameter. The default value is 1024 MB. This parameter should be set to a value small enough to not waste memory unnecessarily (for example, 256 MB)

3.       “yarn.nodemanager.resource.memory-mb” - Determines the maximum amount of memory YARN can use on a node overall. The default value is 8192 MB. Should be set total memory available on node minus memory needed for OS.

Dynamic resource allocation-

Enables applications to release executors temporarily so that other applications can use the

allocated resources. Set “spark.dynamicAllocation.enabled” to true.

You should also enable Spark’s shuffle service, which is used to serve executors’ shuffle files even after the executors are no longer available. If an executor’s shuffle files are requested and the executor isn’t available while the service isn’t enabled, shuffle files will need to be recalculated, which wastes resources. Therefore, you should always enable the shuffle service when enabling dynamic allocation. To enable –

1.       Add spark-<version>-shuffle.jar to classpath of all node mangers.

2.       Set “yarn.nodemanager.aux-services” à { mapreduce_shuffle, spark_shuffle}

3.       Set “yarn.nodemanager.aux-services.spark_shuffle.class” -> “org.apache.spark.network.yarn.YarnShuffleService”

4.       Set “spark.shuffle.service.enabled” -> true

You can control the number of executors with these parameters:

1.       “spark.dynamicAllocation.minExecutors” – Minimum number of executors for your application

2.       “spark.dynamicAllocation.maxExecutors” - Maximum number of executors for your application

3.       “spark.dynamicAllocation.initialExecutors” - Initial number of executors for your application