[MAPREDUCE-1308] reduce tasks stall and are eventually killed

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Cannot Reproduce
• Component/s: tasktracker

Environment

20-node cluster, 8 cores per machine, 32GB memory, Fedora Linux

• Assignee: Unassigned
• Reporter: Brian Karlak
• Created: 17/Dec/09 06:11
• Updated: 29/Jul/14 21:17
• Resolved: 29/Jul/14 21:17

Description

We recently migrated our 0.19.2 cluster from Gentoo Linux to Fedora Linux. Everything was running smoothly before, but now about 5%-10% of our jobs have at least one reduce task that stalls out and is eventually killed with the message:

Task attempt_200912102211_1648_r_000009_0 failed to report status for 6003 seconds. Killing!

The task is then re-launched and completes successfully, usually in a couple of minutes.

This is problematic because our scheduled Hadoop jobs now take an extra hour-and-a-half to run (6000 seconds).

There are no indications in the logs that anything is amiss. The task starts, a small amount of the copy/shuffle runs, and then nothing is else is heard from the task until it is killed. I will attach the relevant parts of the TaskTracker logs in the comments.