[SPARK-12189] UnifiedMemoryManager double counts storage memory freed

• Type: Bug
• Status: Resolved
• Priority: Blocker
• 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: Josh Rosen
• Reporter: Andrew Or
• Created: 08/Dec/15 02:46
• Updated: 10/Dec/15 00:54
• Resolved: 10/Dec/15 00:54

Description

When execution evicts storage, we decrement the storage memory in two places:

(1) https://github.com/apache/spark/blob/3e7e05f5ee763925ed60410d7de04cf36b723de1/core/src/main/scala/org/apache/spark/memory/StorageMemoryPool.scala#L131

(2) https://github.com/apache/spark/blob/3e7e05f5ee763925ed60410d7de04cf36b723de1/core/src/main/scala/org/apache/spark/memory/StorageMemoryPool.scala#L133

(1) calls MemoryStore#ensureFreeSpace, which internally calls MemoryManager#releaseStorageMemory for each block it drops. This call lowers the storage memory used by the block size.

A seemingly simple fix is just to remove the line in (2). However, this bug is actually masked by SPARK-12165, so this one must be fixed after that one. Josh actually has an outstanding patch to fix both: https://github.com/apache/spark/pull/10170

Issue Links

• blocks: SPARK-12165 Execution memory requests may fail to evict storage blocks if storage memory usage is below max memory
• links to: [Github] Pull Request #10170 (JoshRosen)