[HIVE-20951] LLAP: Set Xms to 50% always

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 3.1.1, 4.0.0
• Fix Version/s: 4.0.0-alpha-1
• Component/s: llap
• Assignee: Gopal Vijayaraghavan
• Reporter: Gopal Vijayaraghavan
• Created: 20/Nov/18 20:27
• Updated: 17/Nov/22 08:54
• Resolved: 29/Nov/18 04:23

Description

The lack of GC pauses is killing LLAP containers whenever the significant amount of memory is consumed by the off-heap structures which aren't cleaned up automatically until the GC runs.

There's a java.nio.DirectByteBuffer.Deallocator which runs when the Direct buffers are garbage collected, which actually does the cleanup of the underlying off-heap buffers.

The lack of Garbage collection activity for several hours while responding to queries triggers a build-up of these off-heap structures which end up forcing YARN to kill the process instead.

It is better to hit a GC pause occasionally rather than to lose a node every few hours.