[SYSTEMDS-2147] Performance read of short and wide sparse matrices

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: SystemML 1.1
• Assignee: Matthias Boehm
• Reporter: Matthias Boehm
• Created: 13/Feb/18 07:55
• Updated: 17/Feb/18 04:09
• Resolved: 17/Feb/18 04:09

Description

Since for sparse matrix reads of many columns we synchronize over the first row of a row of blocks, there is unnecessary contention with many read threads and a small number of rows (e.g., 10K rows -> 10 blocks, but 24 threads reading)