[jira] [Commented] (REEF-1002) Add Vortex examples that cause large serialization cost

["Yunseong Lee (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/REEF-1002?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15020863#comment-15020863 ] 

Yunseong Lee commented on REEF-1002:
------------------------------------

There are two examples [~johnyangk] wrote on Vortex in the early days. One is matrix multiplication and the other is logistic regression. We used basic data structures at most (e.g., ArrayList<Vector<Double>>) to minimize serialization cost. But now, we can encapsulate classes to impose more serialization cost in the input, output, and user function.

> Add Vortex examples that cause large serialization cost
> -------------------------------------------------------
>
>                 Key: REEF-1002
>                 URL: https://issues.apache.org/jira/browse/REEF-1002
>             Project: REEF
>          Issue Type: Improvement
>          Components: Vortex
>            Reporter: Yunseong Lee
>
> REEF-504 is for suggesting better serialization in Vortex. But the current examples are too simple to test the serialization cost: AddOne uses Integer, and HelloVortex uses null.
> We may need an example to cause high cost in (de-)serialization of input, output, user function, so we can check how performance is changed when we optimize serialization as suggested in REEF-504.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)