[SPARK-11898] Use broadcast for the global tables in Word2Vec

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.5.2
• Fix Version/s: 2.0.0
• Component/s: MLlib
• Assignee: yuhao yang
• Reporter: yuhao yang
• Created: 21/Nov/15 03:47
• Updated: 08/Jul/16 12:52
• Resolved: 01/Dec/15 09:27

Description

syn0Global and sync1Global in word2vec are quite large objects with size (vocab * vectorSize * 8), yet they are passed to worker using basic task serialization.

Use broadcast can greatly improve the performance. My benchmark shows that, for 1M vocabulary and default vectorSize 100, changing to broadcast can help,
1. decrease the worker memory consumption by 45%.
2. decrease running time by 40%.

Issue Links

• is related to: SPARK-16440 Undeleted broadcast variables in Word2Vec causing OoM for long runs
• links to: [Github] Pull Request #9878 (hhbyyh)