[SPARK-12081] Make unified memory management work with small heaps

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: 1.6.0
• Fix Version/s: 1.6.0
• Component/s: Spark Core
• Target Version/s: 1.6.0
• Assignee: Andrew Or
• Reporter: Andrew Or
• Created: 01/Dec/15 22:36
• Updated: 02/Dec/15 03:51
• Resolved: 02/Dec/15 03:51

Description

By default, Spark drivers and executors are 1GB. With the recent unified memory mode, only 250MB is set aside for non-storage non-execution purposes (spark.memory.fraction is 75%). However, especially in local mode, the driver needs at least ~300MB. Some local jobs started to OOM because of this.

Two mutually exclusive proposals:
(1) First, cut out 300 MB, then take 75% of what remains
(2) Use min(75% of JVM heap size, JVM heap size - 300MB)

Issue Links

• links to: [Github] Pull Request #10081 (andrewor14)