Редактиране

Споделяне чрез

Optimize Apache Spark applications in HDInsight

  • Статия

This article provides an overview of strategies to optimize Apache Spark applications on Azure HDInsight.

Overview

You might face below common Scenarios

  • The same spark job is slower than before in the same HDInsight cluster
  • The spark job is slower in HDInsight cluster than on-premise or other third party service provider
  • The spark job is slower in one HDI cluster than another HDI cluster

The performance of your Apache Spark jobs depends on multiple factors. These performance factors include:

  • How your data is stored
  • How the cluster is configured
  • The operations that are used when processing the data.
  • Unhealthy yarn service
  • Memory constraints due to improperly sized executors and OutOfMemoryError
  • Too many tasks or too few tasks
  • Data skew caused a few heavy tasks or slow tasks
  • Tasks slower in bad nodes

Step 1: Check if your yarn service is healthy

  1. Go to Ambari UI:
  • Check if ResourceManager or NodeManager alerts
  • Check ResourceManager and NodeManager status in YARN > SUMMARY: All NodeManager should be in Started and only Active ResourceManager should be in Started

  1. Check if Yarn UI is accessible through https://YOURCLUSTERNAME.azurehdinsight.net/yarnui/hn/cluster

  2. Check if any exceptions or errors in ResourceManager log in /var/log/hadoop-yarn/yarn/hadoop-yarn-resourcemanager-*.log

See more information in Yarn Common Issues

Step 2: Compare your new application resources with yarn available resources

  1. Go to Ambari UI > YARN > SUMMARY, check CLUSTER MEMORY in ServiceMetrics

  2. Check yarn queue metrics in details:

  • Go to Yarn UI, check Yarn scheduler metrics through https://YOURCLUSTERNAME.azurehdinsight.net/yarnui/hn/cluster/scheduler
  • Alternatively, you can check yarn scheduler metrics through Yarn Rest API. For example, curl -u "xxxx" -sS -G "https://YOURCLUSTERNAME.azurehdinsight.net/ws/v1/cluster/scheduler". For ESP, you should use domain admin user.
  1. Calculate total resources for your new application
  • All executors resources: spark.executor.instances * (spark.executor.memory + spark.yarn.executor.memoryOverhead) and spark.executor.instances * spark.executor.cores. See more information in spark executors configuration
  • ApplicationMaster
    • In cluster mode, use spark.driver.memory and spark.driver.cores
    • In client mode, use spark.yarn.am.memory+spark.yarn.am.memoryOverhead and spark.yarn.am.cores

Note

yarn.scheduler.minimum-allocation-mb <= spark.executor.memory+spark.yarn.executor.memoryOverhead <= yarn.scheduler.maximum-allocation-mb

  1. Compare your new application total resources with yarn available resources in your specified queue

Step 3: Track your spark application

  1. Monitor your running spark application through Spark UI

  2. Monitor your complete or incomplete spark application through Spark History Server UI

We need to identify below symptoms through Spark UI or Spark History UI:

  • Which stage is slow
  • Are total executor CPU v-cores fully utilized in Event-Timeline in Stage tab
  • If using spark sql, what's the physical plan in SQL tab
  • Is DAG too long in one stage
  • Observe tasks metrics(input size, shuffle write size, GC Time) in Stage tab

See more information in Monitoring your Spark Applications

Step 4: Optimize your spark application

There are many optimizations that can help you overcome these challenges, such as caching, and allowing for data skew.

In each of the following articles, you can find information on different aspects of Spark optimization.

Optimize Spark SQL partitions

  • spark.sql.shuffle.paritions is 200 by default. We can adjust based on the business needs when shuffling data for joins or aggregations.
  • spark.sql.files.maxPartitionBytes is 1G by default in HDI. The maximum number of bytes to pack into a single partition when reading files. This configuration is effective only when using file-based sources such as Parquet, JSON and ORC.
  • AQE in Spark 3.0. See Adaptive Query Execution

Next steps