Package org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources

Class DefaultOOMHandler

java.lang.Object

org.apache.hadoop.yarn.server.nodemanager.containermanager.linux.resources.DefaultOOMHandler

All Implemented Interfaces:

Runnable

@Public
@Evolving
public class DefaultOOMHandler
extends Object
implements Runnable

A very basic OOM handler implementation. See the javadoc on the run() method for details.

Field Summary

Fields

Modifier and Type	Field	Description
protected static final org.slf4j.Logger	LOG

Constructor Summary

Constructors

Constructor	Description
DefaultOOMHandler(Context context, boolean enforceVirtualMemory)	Create an OOM handler.

Method Summary

Modifier and Type	Method	Description
protected CGroupsHandler	getCGroupsHandler()
protected boolean	killContainer()	Choose and kill a container in case of OOM.
void	run()	It is called when the node is under an OOM condition.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Details

LOG

protected static final org.slf4j.Logger LOG

Constructor Details

DefaultOOMHandler

public DefaultOOMHandler(Context context,
 boolean enforceVirtualMemory)

Create an OOM handler. This has to be public to be able to construct through reflection.

Parameters:

context - node manager context to work with

enforceVirtualMemory - true if virtual memory needs to be checked, false if physical memory needs to be checked instead

Method Details

getCGroupsHandler

@VisibleForTesting
protected CGroupsHandler getCGroupsHandler()

run

public void run()

It is called when the node is under an OOM condition. All processes in all sub-cgroups are suspended. We need to act fast, so that we do not affect the overall system utilization. In general we try to find a newly launched container that exceeded its limits. The justification is cost, since probably this is the one that has accumulated the least amount of uncommitted data so far. OPPORTUNISTIC containers are always killed before any GUARANTEED containers are considered. We continue the process until the OOM is resolved.

Specified by:

run in interface Runnable

killContainer

protected boolean killContainer()

Choose and kill a container in case of OOM. We try to find the most recently launched OPPORTUNISTIC container that exceeds its limit and fall back to the most recently launched OPPORTUNISTIC container If there is no such container found, we choose to kill a GUARANTEED container in the same way.

Returns:

true if a container is killed, false otherwise