[SYSTEMDS-1881] Tuning parfor degree of parallelism for operations

• Type: Task
• Status: Closed
• Priority: Major
• Resolution: Done
• Fix Version/s: SystemML 0.15
• Assignee: Matthias Boehm
• Reporter: Matthias Boehm
• Created: 02/Sep/17 06:29
• Updated: 09/Sep/17 04:02
• Resolved: 02/Sep/17 23:32

Description

Currently, we assign remaining parfor parallelism conservatively to operations of the parfor body. Consider, for example, a Kmeans or MSVM scenario with 10 runs or 10 classes respectively. On a box with 16 HW threads, we assign k=10 to the parfor and floor(16/10) to remaining operations. Since it is usually a good idea to slightly over-provision CPU in order to get full utilization (due to barriers at the end of each operation), we should tune this to round(16/10) which provides performance improvements of about 15% in above examples.