[SPARK-12869] Optimize conversion from BlockMatrix to IndexedRowMatrix

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 1.6.0
• Fix Version/s: 2.0.0
• Component/s: MLlib
• Target Version/s: 2.0.0
• Assignee: Fokko Driesprong
• Reporter: Fokko Driesprong
• Shepherd: Xiangrui Meng
• Created: 17/Jan/16 22:09
• Updated: 26/Jan/17 13:35
• Resolved: 15/Apr/16 00:32

Description

In the current implementation of the BlockMatrix, the conversion to the IndexedRowMatrix is done by converting it to a CoordinateMatrix first. This is somewhat ok when the matrix is very sparse, but for dense matrices this is very inefficient.

Issue Links

• relates to: SPARK-19368 Very bad performance in BlockMatrix.toIndexedRowMatrix()
• links to: [Github] Pull Request #10839 (Fokko)