[MAPREDUCE-3769] [Gridmix] Improve the way job monitor maintains running jobs

• Type: Improvement
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 3.0.0-alpha1
• Component/s: contrib/gridmix
• Labels: gridmix job-monitor
• Tags: gridmix job monitor
• Assignee: Unassigned
• Reporter: Amar Kamat
• Created: 31/Jan/12 05:22
• Updated: 12/May/16 18:23

Description

Gridmix maintains a list (L) of running jobs via JobMonitor. As soon as a job is submitted, a handle for that job is cached inside the JobMonitor. The JobMonitor does the following in a thread:

1. remove the first job in the list, say j
2. if j is complete :
     goto #1.
   else :
     add j to the end of the list L. 
     sleep for 5 seconds. 
     goto #1.

Gridmix STRESS mode logic uses the list L to compute the cluster load. It iterates over map/reduce progress of each and every job in L to figure out the pending+running task count. We need to investigate and optimize the JobMonitor algorithm and make sure that the total number of completed jobs in L is minimum. The overhead of polling for the map and reduce task progress of a completed job is pretty high as it incurs an additional (RPC) step of contacting the JobHistory server.

Issue Links

• is related to: MAPREDUCE-1687 Stress submission policy does not always stress the cluster.
• is related to: MAPREDUCE-3481 [Gridmix] Improve STRESS mode locking
• is related to: MAPREDUCE-3787 [Gridmix] Improve STRESS mode