YARN not preempting resources based on fair shares when running a Spark job

问题

I have a problem with re-balancing Apache Spark jobs resources on YARN Fair Scheduled queues.

For the tests I've configured Hadoop 2.6 (tried 2.7 also) to run in pseudo-distributed mode with local HDFS on MacOS. For job submission used "Pre-build Spark 1.4 for Hadoop 2.6 and later" (tried 1.5 also) distribution from Spark's website.

When tested with basic configuration on Hadoop MapReduce jobs, Fair Scheduler works as expected: When resources of the cluster exceed some maximum, fair shares are calculated and resources for jobs in different queues are preempted and balanced based on these calculations.

The same test is ran with Spark jobs, in that case YARN is making correct calculations of the fair shares for each job, but resources for Spark containers are not re-balanced.

Here are my conf files:

$HADOOP_HOME/etc/hadoop/yarn-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
<configuration>
   <property>
      <name>yarn.nodemanager.aux-services</name>
      <value>mapreduce_shuffle</value>
   </property>
   <property>
      <name>yarn.nodemanager.aux-services.spark_shuffle.class</name>
      <value>org.apache.spark.network.yarn.YarnShuffleService</value>
   </property>
   <property>
      <name>yarn.resourcemanager.scheduler.class</name>
      <value>org.apache.hadoop.yarn.server.resourcemanager.scheduler.fair.FairScheduler</value>
   </property>
   <property>
      <name>yarn.scheduler.fair.preemption</name>
      <value>true</value>
   </property>
</configuration>

$HADOOP_HOME/etc/hadoop/fair-scheduler.xml

<?xml version="1.0" encoding="UTF-8"?>
<allocations>
   <defaultQueueSchedulingPolicy>fair</defaultQueueSchedulingPolicy>
   <queue name="prod">
      <weight>40</weight>
      <schedulingPolicy>fifo</schedulingPolicy>
   </queue>
   <queue name="dev">
      <weight>60</weight>
      <queue name="eng" />
      <queue name="science" />
   </queue>
   <queuePlacementPolicy>
      <rule name="specified" create="false" />
      <rule name="primaryGroup" create="false" />
      <rule name="default" queue="dev.eng" />
   </queuePlacementPolicy>
</allocations>

$HADOOP_HOME/etc/hadoop/core-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
<configuration>
   <property>
      <name>fs.defaultFS</name>
      <value>hdfs://localhost:9000</value>
   </property>
</configuration>

$HADOOP_HOME/etc/hadoop/core-site.xml

<?xml version="1.0" encoding="UTF-8"?>
<?xml-stylesheet type="text/xsl" href="configuration.xsl"?>
<configuration>
   <property>
      <name>dfs.replication</name>
      <value>1</value>
   </property>
</configuration>

And the test case is:

Run a job on the "prod" queue with weight 40 (must allocate 40% of all resources), as expected the job takes all needed free resources (62,5% of the clusters resources).

./bin/spark-submit --class org.apache.spark.examples.SparkPi \
--master yarn-cluster \
--driver-memory 512M \
--executor-memory 768M \
--executor-cores 1 \
--num-executors 2 \
--queue prod \
lib/spark-examples*.jar 100000

After that run the same job on the "dev.eng" queue with weight 60, that mean the job must allocate 60% of all resources and decrease the first job's resources to ~40%.

./bin/spark-submit --class org.apache.spark.examples.SparkPi \
--master yarn-cluster \
--driver-memory 512M \
--executor-memory 768M \
--executor-cores 1 \
--num-executors 2 \
--queue dev.eng \
lib/spark-examples*.jar 100000

Unfortunately, cluster resources are not changing - 62,5% for the first job and 37,5% for second.

回答1:

You need to set one of the preemption timeouts in your allocation xml. One for minimum share and one for fair share, both are in seconds. By default, the timeouts are not set.

From Hadoop: The Definitive Guide 4th Edition

If a queue waits for as long as its minimum share preemption timeout without receiving its minimum guaranteed share, then the scheduler may preempt other containers. The default timeout is set for all queues via the defaultMinSharePreemptionTimeout top-level element in the allocation file, and on a per-queue basis by setting the minSharePreemptionTimeout element for a queue.

Likewise, if a queue remains below half of its fair share for as long as the fair share preemption timeout, then the scheduler may preempt other containers. The default timeout is set for all queues via the defaultFairSharePreemptionTimeout top-level element in the allocation file, and on a per-queue basis by setting fairSharePreemptionTimeout on a queue. The threshold may also be changed from its default of 0.5 by setting defaultFairSharePreemptionThreshold and fairSharePreemptionThreshold (per-queue).

回答2:

Fair Scheduler doesn't kill containers for the first job, It only wait until some resources are free up and reserve them to be used by the second job. If not resources are free up from the first job, the scheduler can not assign those resources to the second job.

In MapReduce jobs, each map or reduce task require to instantiate a new container and the scheduler can block the job to instantiate new containers if it has exceeded its quote (based on the queue capacity).

In Spark the things are different, the executors are being initiated at the beginning of the job and the different tasks (stages) are sent to them. Then the resources are not being free up and they can not be reallocated.

May be dynamic allocation could help: http://spark.apache.org/docs/1.6.1/configuration.html#dynamic-allocation

来源：https://stackoverflow.com/questions/33082088/yarn-not-preempting-resources-based-on-fair-shares-when-running-a-spark-job

标签

Hadoop

apache-spark

yarn

job-scheduling