RM and JHS volume management

Describes the RM volume sharding and separating Resource Manager (RM) and Job History Server (JHS) volume.

Overview

If all the data of YARN Resource Manager (RM) is stored in a single data volume:
  • The amount of data stored increases over a time, and the data volume becomes too large.
  • The RM has to write to the same volume multiple times, for every application. This slows down the application response, and affects the user experience.

Data sharding overcomes this limitation by splitting the large volume of data in to smaller chunks, called shards, and by storing the data across multiple nodes. Data sharding helps to improve the response time, and to scale the RM operations efficiently.

Prerequisites for Sharding the RM Volume

IMPORTANT
You must either separate the RM and JHS volumes together with RM volume sharding, or separate the RM and JHS volumes without using RM volume sharding. Using only RM volume sharding is not supported.
  • Starting from EEP 8.1.0 or higher, this feature is available by default.
  • Cluster must not have any jobs running before separating a volume for Job History Server (JHS).

Sharding the RM Volume

By Default, both Resource Manager (RM) and Job History Server (JHS) services are stored in the same volume at MapR FS as follows:
volumename - mapr.resourcemanager.volume
mountdir - /var/mapr/cluster/yarn
To check the EEP version, see Checking the EEP Version
Seperating RM and JHS:
NOTE
By default, RM and JHS volume are not configured as seperate volumes.

  1. To seperate RM and JHS volume, run the following command:

    yarn.resourcemanager.dir.new-volume-path-support.enabled - true - to yarn-site.xml
yarn.app.mapreduce.am.staging-dir - /var/mapr/cluster/hs - to mapred-site.xml
  2. Restart the RM and JHS. After successful restart, new volumes will be created and old data will be moved to separate volume as shown in the following example:
    For ResourceManager:
  mountdir - /var/mapr/cluster/yarn/rm
  volumename - mapr.resourcemanager.volume
For JobHistoryServer:
  mountdir - /var/mapr/cluster/hs
  volumename - mapr.historyserver.volume
Enable RM sharding:
NOTE
By default, RM volume sharding is not configured.
To enable RM sharding, In yarn-site.xml file, set the yarn.resourcemanager.dir.volume-sharding.enabled to true, for example:
yarn.resourcemanager.dir.volume-sharding.enabled - true
yarn.resourcemanager.dir.volume-count - number of ResorceManager volumes (default 4)
Default value of number of shards is four, and you can increase or decrease this number as necessary.