[jira] [Created] (SYSTEMML-1968) Improve cost-based optimizer for distributed operations

["Matthias Boehm (JIRA)" <ji...@apache.org>]

Matthias Boehm created SYSTEMML-1968:
----------------------------------------

             Summary: Improve cost-based optimizer for distributed operations
                 Key: SYSTEMML-1968
                 URL: https://issues.apache.org/jira/browse/SYSTEMML-1968
             Project: SystemML
          Issue Type: Sub-task
            Reporter: Matthias Boehm


The current cost-based codegen optimizer has a couple of short-comings that lead to missed opportunities for large-scale computation:

* Cost model unaware of distributed operations, which sometimes leads to unnecessary fusion of cell-wise vector operations into distributed row operations, which hurts performance due to additional side inputs that needs to be broadcast to all executors.
* The current identification of interesting points ignores common Cell entries. In order to reason about above problematic, we need to consider cell to row switches, i.e., Cell templates that are merged into row templates as well.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)