[MAPREDUCE-5946] Last spill of map task is not necessary for final merge

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 2.4.0
• Component/s: performance, task
• Labels: newbie performance
• Assignee: jaehoon ko
• Reporter: jaehoon ko
• Created: 27/Jun/14 00:44
• Updated: 27/Jun/14 01:16

Description

In map task, merge starts only after the last spill is completely written to disk. This is not necessary nor efficient because the last spill should to be reloaded soon for merge, probably immediately because spills are merged in the order of their sizes and the last spill is likely smallest. OS page cache is not the answer due to its opportunistic nature.

I'm starting to work on this. Please give me your thoughts.