[SPARK-12253] UnifiedMemoryManager race condition: storage can starve new tasks

• 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: Andrew Or
• Reporter: Andrew Or
• Created: 10/Dec/15 02:54
• Updated: 10/Dec/15 23:30
• Resolved: 10/Dec/15 23:30

Description

The following race condition is possible with the existing code in unified memory management:

(1) Existing tasks collectively occupy all execution memory
(2) New task comes in and blocks while existing tasks spill
(3) After tasks finish spilling, another task jumps in and puts in a large block, stealing the freed memory
(4) New task still cannot acquire memory and goes back to sleep

Issue Links

• links to: [Github] Pull Request #10240 (andrewor14)