Optimizing Node Configuration

Scenario

After the scheduler of a big data cluster is properly configured, you can adjust the available memory, CPU resources, and local disk of each node to optimize the performance.

The configuration items are as follows:

• Available memory
• Number of vCPUs
• Physical CPU usage
• Coordination of memory and CPU resources
• Local disk

Procedure

For details about how to adjust parameter settings, see Modifying Cluster Service Configuration Parameters.

• Available memory

  Except the memory allocated to the OS and other services, allocate as much as possible memory to Yarn. You can adjust the following parameters to improve resource utilization.

  Assume that a container uses 512 MB memory by default, then the memory usage formula is: 512 MB x Number of containers.

  By default, the Map or Reduce container uses one vCPU and 1,024 MB memory, and ApplicationMaster uses 1,536 MB memory.

  Parameter	Description	Default Value
  yarn.nodemanager.resource.memory-mb	Physical memory that can be allocated to containers, in MB. The value must be greater than 0. You are advised to set the parameter value to 75% to 90% of the total physical memory of nodes. If the node has permanent processes of other services, reduce this parameter value to reserve sufficient resources for the processes. If the total physical memory space of a node is large and there is no resident process of other services, set this parameter to the total physical memory minus the memory occupied by the resident processes of NodeManager.	16384

• Number of vCPUs

  You are advised to set this parameter to 1.5 to 2 times the number of logical CPUs. If the upper layer computing applications have low computing capability requirements, you can set the parameter to two times the number of logical CPUs.

  Parameter	Description	Default Value
  yarn.nodemanager.resource.cpu-vcores	Number of vCPUs that can be used by Yarn on the node. The default value is 8. You are advised to set the value to 1.5 to 2 times the number of logical CPUs. If the task is computing-intensive, set this parameter to the number of logical CPU cores. If the task is not computing-intensive, set this parameter to 1.5 to 2 times the number of logical CPU cores. If the number of CPU cores used by a task differs greatly from the memory resources, configure the CPU resources based on the memory resources. For example, most tasks use one core and 3 GB memory. If yarn.nodemanager.resource.memory-mb is 380 GB, set this parameter to 128.	8

• Physical CPU usage

  You are advised to reserve appropriate CPUs for the OS and the processes, such as database and HBase, and allocate the remaining CPUs to Yarn. You can set the following parameters to adjust the physical CPU usage.

  Parameter	Description	Default Value
  yarn.nodemanager.resource.percentage-physical-cpu-limit	Physical CPU percentage that can be used by Yarn on a node. The default value is 90, indicating that no CPU control is implemented and Yarn can use all CPU resources. You can only view the parameter. To change the value of this parameter, set the value of RES_CPUSET_PERCENTAGE of YARN. You are advised to set this parameter to the percentage of CPU resources that can be used by the YARN cluster. For example, If 20% of CPU resources are used by other services (such as HBase, HDFS, and Hive) and system processes on the node, the CPU resources can be scheduled for Yarn is 1 - 20% = 80%. Therefore, you can set this parameter to 80.	90

• Local disk

  MapReduce writes the intermediate job execution results in local disks. Therefore, configure disks as much as possible and disk space as large as possible. A simple way is to configure the same number of disks as DataNode except for the last directory.

  

  Use commas (,) to separate multiple disks.

  Parameter	Description	Default Value
  yarn.nodemanager.log-dirs	Directories in which logs are stored. Multiple directories can be specified. Storage location of container logs. The default value is %{@auto.detect.datapart.nm.logs}. If there is a data partition, a path list similar to /srv/BigData/hadoop/data1/nm/containerlogs,/srv/BigData/hadoop/data2/nm/containerlogs is generated based on the data partition. If there is no data partition, the default path /srv/BigData/yarn/data1/nm/containerlogs is generated. In addition to using expressions, you can enter a complete list of paths, such as /srv/BigData/yarn/data1/nm/containerlogs or /srv/BigData/yarn/data1/nm/containerlogs,/srv/BigData/yarn/data2/nm/containerlogs. In this way, data is stored in all the configured directories, which are usually on different devices. To ensure disk I/O load balancing, you are advised to provide several paths and each path corresponds to an independent disk. The localized log directory of the application exists in the relative path /application_%{appid}. The log directory of an independent container, that is, container_{$contid}, is the subdirectory of this directory. Each container directory contains the stderr, stdin, and syslog files generated by the container. To add a directory, for example, /srv/BigData/yarn/data2/nm/containerlogs, you need to delete the files in /srv/BigData/yarn/data2/nm/containerlogs first. Then, assign the same read and write permissions to /srv/BigData/yarn/data2/nm/containerlogs as those of /srv/BigData/yarn/data1/nm/containerlogs, and change /srv/BigData/yarn/data1/nm/containerlogs to /srv/BigData/yarn/data1/nm/containerlogs,/srv/BigData/yarn/data2/nm/containerlogs. You can add directories, but do not modify or delete existing directories. Otherwise, NodeManager data will be lost and services will be unavailable. Default value: %{@auto.detect.datapart.nm.logs} Exercise caution when modifying this parameter. If the configuration is incorrect, the services are unavailable. If the value of this configuration item at the role level is changed, the value of this configuration item at all instance levels will be changed. If the value of this configuration item at the instance level is changed, the value of this configuration item of other instances remains unchanged.	%{@auto.detect.datapart.nm.logs}
  yarn.nodemanager.local-dirs	Storage location of files after localization. The default value is %{@auto.detect.datapart.nm.localdir}. If there is a data partition, a path list similar to /srv/BigData/hadoop/data1/nm/localdir,/srv/BigData/hadoop/data2/nm/localdir is generated based on the data partition. If there is no data partition, the default path /srv/BigData/yarn/data1/nm/localdir is generated. In addition to using expressions, you can enter a complete list of paths, such as /srv/BigData/yarn/data1/nm/localdir or /srv/BigData/yarn/data1/nm/localdir,/srv/BigData/yarn/data2/nm/localdir. In this way, data is stored in all the configured directories, which are usually on different devices. To ensure disk I/O load balancing, you are advised to provide several paths and each path corresponds to an independent disk. The localized file directory of the application is stored in the relative path /usercache/%{user}/appcache/application_%{appid}. The working directory of an independent container, that is, container_%{contid}, is the subdirectory of the directory. To add a directory, for example, /srv/BigData/yarn/data2/nm/localdir, you need to delete the files in /srv/BigData/yarn/data2/nm/localdir first. Then, assign the same read and write permissions to /srv/BigData/hadoop/data2/nm/localdir as those of /srv/BigData/hadoop/data1/nm/localdir, and change /srv/BigData/yarn/data1/nm/localdir to /srv/BigData/yarn/data1/nm/localdir,/srv/BigData/yarn/data2/nm/localdir. You can add directories, but do not modify or delete existing directories. Otherwise, NodeManager data will be lost and services will be unavailable. Default value: %{@auto.detect.datapart.nm.localdir} Exercise caution when modifying this parameter. If the configuration is incorrect, the services are unavailable. If the value of this configuration item at the role level is changed, the value of this configuration item at all instance levels will be changed. If the value of this configuration item at the instance level is changed, the value of this configuration item of other instances remains unchanged.	%{@auto.detect.datapart.nm.localdir}