Managing Resources with YARN

YARN (Yet Another Resource Negotiator) can manage Hadoop applications like MapReduce so that applications can reserve resources like CPU and memory so that resources are not denied to other applications. YARN applications request resources from a resource manager.

If YARN is already used on the cluster, then you can configure SAS LASR Analytic Server instances and high-performance procedures to participate in the resource accounting that YARN performs. This enables administrators to have a complete view of resource usage. Otherwise, having a mix of some applications accounting for their resources with YARN and others that are not essentially results in no management at all.

Note: Some YARN implementations create CGroups automatically, which can interfere with CGroups that you administer manually.

In order to integrate with YARN, the following settings in the resource.settings file are used.

# The number of cores to allocate to each host's container.
export TKMPI_YARN_CORES=1
 
# The amount of memory in megabytes to reserve.
export TKMPI_MEMSIZE=30000
 
# The priority of the application if scheduler uses it.
export TKMPI_YARN_PRIORITY=2
 
# Length of time TKGrid should wait for the resource reservation in seconds.
export TKMPI_YARN_TIMEOUT=3600
 
# The queue to submit the job to.
export TKMPI_YARN_QUEUE=default
 
# The next setting must be on one line in the resource.settings file, 
# but is split for readability.
export TKMPI_RESOURCEMANAGER="java -cp \"`$HADOOP_HOME/bin/hadoop 
 classpath`\" com.sas.grid.provider.yarn.tkgrid.JobLauncher 
 --masterMem 2000 --javaMem 500 --hostlist \$TKMPI_YARN_HOSTS \
 --cores \$TKMPI_YARN_CORES --memory \$TKMPI_MEMSIZE \
 --priority \$TKMPI_YARN_PRIORITY --timeout \$TKMPI_YARN_TIMEOUT
 --jobname $TKMPI_APPNAME --queue $TKMPI_YARN_QUEUE"

Note: The TKMPI_YARN_HOSTS and TKMPI_APPNAME variables are automatically set by SAS software.

The TKMPI_MEMSIZE variable specifies the amount of memory for YARN to reserve. SAS also uses the value to self-govern the memory allocations that are performed by the server or high-performance procedure.

YARN must be configured on the cluster. The JobLauncher class starts a YARN application to request the specified resources on each machine so that YARN knows it cannot allocate those resources to other applications. If YARN does not grant the resource in the time-out period (TKMPI_YARN_TIMEOUT), then the initializing of the server or job fails.

The servers and jobs are accounted for as YARN applications on the ResourceManager web user interface.

Hadoop ResourceManager Web User Interface

If the State field for an application indicates RUNNING and the Progress percentage is at 50%, then the resources are reserved and the application is running. This is the normal state.

If the State field indicates ACCEPTED and the Tracking UI field indicates UNASSIGNED, then the application is not running. Check if you are near or at capacity.