On an Ambari cluster, you can enable CPU Scheduling to enable cgroups. On a non-Ambari
cluster, you must configure certain properties in
yarn-site.xml on the
ResourceManager and NodeManager hosts to enable cgroups.
cgroups is a Linux kernel feature. cgroups is supported on the following Linux operating systems:
CentOS 6.9, 7.3
RHEL 6.9, 7.3
At this time there is no cgroups equivalent for Windows. Cgroups are not enabled by default on HDP. Cgroups require that the HDP cluster be Kerberos enabled.
The following commands must be run on every reboot of the
NodeManager hosts to set up the CGroup hierarchy. Note that operating systems use
different mount points for the CGroup interface. Replace
/sys/fs/cgroup with your operating system equivalent.
mkdir -p /sys/fs/cgroup/cpu/hadoop-yarn chown -R yarn /sys/fs/cgroup/cpu/hadoop-yarn mkdir -p /sys/fs/cgroup/memory/hadoop-yarn chown -R yarn /sys/fs/cgroup/memory/hadoop-yarn mkdir -p /sys/fs/cgroup/blkio/hadoop-yarn chown -R yarn /sys/fs/cgroup/blkio/hadoop-yarn mkdir -p /sys/fs/cgroup/net_cls/hadoop-yarn chown -R yarn /sys/fs/cgroup/net_cls/hadoop-yarn mkdir -p /sys/fs/cgroup/devices/hadoop-yarn chown -R yarn /sys/fs/cgroup/devices/hadoop-yarn
- To enable cgroups on an Ambari cluster, select YARN > Configs on the Ambari dashboard, then click CPU Isolation under CPU. Click Save, then restart all cluster components that require a restart. cgroups should be enabled along with CPU Scheduling.
On a non-Ambari cluster, set the following properties in the
/etc/hadoop/conf/yarn-site.xmlfile on the ResourceManager and NodeManager hosts.
<property> <name>yarn.nodemanager.container-executor.class</name> <value>org.apache.hadoop.yarn.server.nodemanager.LinuxContainerExecutor</value> </property>
<property> <name>yarn.nodemanager.linux-container-executor.group</name> <value>hadoop</value> </property>
<property> <name> yarn.nodemanager.linux-container-executor.resources-handler.class</name> <value> org.apache.hadoop.yarn.server.nodemanager.util.cgroupsLCEResourcesHandler </value> </property>
<property> <name>yarn.nodemanager.linux-container-executor.cgroups.hierarchy</name> <value>/hadoop-yarn</value> </property>
<property> <name>yarn.nodemanager.linux-container-executor.cgroups.mount</name> <value>false</value> </property>
<property> <name>yarn.nodemanager.linux-container-executor.cgroups.mount-path</name> <value>/sys/fs/cgroup</value> </property>
Set the Percentage of CPU used by YARN
Set the percentage of CPU that can be allocated for YARN containers. In most cases, the default value of 100% should be used. If you have another process that needs to run on a node that also requires CPU resources, you can lower the percentage of CPU allocated to YARN to free up resources for the other process.
<property> <name>yarn.nodemanager.resource.percentage-physical-cpu-limit</name> <value>100</value> </property>
Set Flexible or Strict CPU limits
CPU jobs are constrained with CPU scheduling and cgroups enabled, but by default these are flexible limits. If spare CPU cycles are available, containers are allowed to exceed the CPU limits set for them. With flexible limits, the amount of CPU resources available for containers to use can vary based on cluster usage -- the amount of CPU available in the cluster at any given time.
You can use cgroups to set strict limits on CPU usage. When strict limits are enabled, each process receives only the amount of CPU resources it requests. With strict limits, a CPU process will receive the same amount of cluster resources every time it runs.
Strict limits are not enabled (set to
false) by default.
<property> <name>yarn.nodemanager.linux-container-executor.cgroups.strict-resource-usage</name> <value>false</value> </property>Note
Irrespective of whether this property is
false, at no point will total container CPU usage exceed the limit set in
yarn.nodemanager.resource.percentage-physical-cpu-limit.NoteCPU resource isolation leverages advanced features in the Linux kernel. At this time, setting
trueis not recommended due to known kernel panics. In addition, with some kernels, setting
yarn.nodemanager.resource.percentage-physical-cpu-limitto a value less than 100 can result in kernel panics. If you require either of these features, you must perform scale testing to determine if the in-use kernel and workloads are stable. As a starting point, Linux kernel version 4.8.1 works with these features. However, testing the features with the desired workloads is very important.