[
https://issues.apache.org/jira/browse/TAJO-223?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13783716#comment-13783716
]
Hyunsik Choi commented on TAJO-223:
-----------------------------------
Hello Keuntae,
Your proposal looks reasonable. If you want to take this issue, please feel
free to ask me.
> Maximize disk read bandwidth utilization of StorageManagerV2 by moving Tuple
> creation role to next()
> ----------------------------------------------------------------------------------------------------
>
> Key: TAJO-223
> URL: https://issues.apache.org/jira/browse/TAJO-223
> Project: Tajo
> Issue Type: Improvement
> Components: storage
> Reporter: Keuntae Park
> Labels: performance
>
> Currently, Tuple creation mechanism of StorageManagerV2 is as follows:
> 1) At file scan, scheduled scanner reads data from disk, makes a Tuple, and
> insert it to the Tuple pool
> 2) next() of the scanner just pulls an already created Tuple from the Tuple
> pool asynchronously
> Because of Tuple creation time, scanner cannot fully use its time to read
> disk, which results in less disk read bandwidth utilization
> So, if Tuple creation role is moved to next() and scanners spend their whole
> time to read file at file scan,
> we can fully utilize disk read bandwidth
--
This message was sent by Atlassian JIRA
(v6.1#6144)
