[
https://issues.apache.org/jira/browse/TAJO-907?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14124897#comment-14124897
]
ASF GitHub Bot commented on TAJO-907:
-------------------------------------
Github user hyunsik commented on the pull request:
https://github.com/apache/tajo/pull/133#issuecomment-54747093
As I mentioned, I improved ExternalSort to make use of OffHeapRowBlock. I'd
like to share a brief benchmark test result.
Experimental environment was as follows:
* Linux,
* Intel Xeon,
* 1 JVM worker, and 4 worker currency.
* A local cluster
Data set is lineitem table of TPC-H 1GB.
The test query is
```
select * from lineitem order by l_orderkey, l_partkey desc;
```
Current implementation: 55 - 60 seconds
After OffHeapRowBlock : 22 - 25 seconds
Roughly, it reduces about 50% query response time. This change only affects
physical operators, so I think that a local cluster benchmark is sufficient.
Later, I'll adopt OffHeapRowBlock to other operators and our operator model
sequentially.
> Implement off-heap tuple block and zero-copy tuple
> --------------------------------------------------
>
> Key: TAJO-907
> URL: https://issues.apache.org/jira/browse/TAJO-907
> Project: Tajo
> Issue Type: Sub-task
> Components: storage
> Reporter: Hyunsik Choi
> Assignee: Hyunsik Choi
> Fix For: 0.9.0
>
> Attachments: TAJO-907.patch, TAJO-907_2.patch
>
>
> The main purpose of this issue is to implement new Tuple structure using
> direct memory.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
