[YARN-284] YARN capacity scheduler doesn't spread MR tasks evenly on an underutilized cluster

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Implemented
• Affects Version/s: 2.0.0-alpha
• Component/s: scheduler
• Assignee: Unassigned
• Reporter: Todd Lipcon
• Created: 25/Oct/11 20:39
• Updated: 22/Dec/12 05:50
• Resolved: 22/Dec/12 00:42

Description

The fair scheduler in MR1 has the behavior that, if a job is submitted to an under-utilized cluster and the cluster has more open slots than tasks in the job, the tasks are spread evenly throughout the cluster. This improves job latency since more spindles and NICs are utilized to complete the job. In MR2 I see this issue causing significantly longer job runtimes when there is excess capacity in the cluster – especially on reducers which sometimmes end up clumping together on a smaller set of nodes which then become disk/network constrained.

Issue Links

• is superceded by: YARN-80 Support delay scheduling for node locality in MR2's capacity scheduler