In a MAP-reduce job, the application master decides how many tasks need to be created for the job to be executed. The number of mapper tasks created is equal to the number of SPLITS. These tasks are usually launched in different JVMs than the application master (governed by data locality).

However, if the job is small, the application master may decide to run the tasks in the same JVM as itself. In such a CASE the overhead of allocating containers for new tasks and monitoring them to gain that would be had in running the tasks in parallel compared to running the tasks sequentially on the same node. Such a job is called an Uber task.

So how does the application master determine if the job is small enough to be run as an uber task? By default, if the job requires LESS than 10 mappers for its processing and one 1 reducer, and input size is less than the size of one HDFS block, then the application master may consider launching the job as an uber task.

If the job doesn’t qualify to be run as an uber task then the app master requests for containers to be launched for all map and reduce tasks from the resource manager.