[jira] [Commented] (HIVE-7144) GC pressure during ORC StringDictionary writes

["Gopal V (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-7144?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14039731#comment-14039731 ] 

Gopal V commented on HIVE-7144:
-------------------------------

Benchmark insert of TPC-H 1Tb scale data

|| Table || Before || After || Diff ||
| part | 46.497 | 45.508 | 0.989 |
| partsupp | 145.031 | 144.841 | 0.19 |
| customer | 55.315 | 55.745 | -0.430 |
| orders | 246.692 | 217.834 | 28.858 |
| lineitem | 959.995 | 875.659 | 84.336 |

This makes negligible difference to the smaller tables, but gives a ~10% boost for orders (50.8Gb ORC) and lineitem (224.5Gb ORC).

The optimization does not benefit the sorted string columns, because for every row read there's a data-copy to populate either the minimum or maximum in the column statistics section (Strings are immutable, Text is Writable).

> GC pressure during ORC StringDictionary writes 
> -----------------------------------------------
>
>                 Key: HIVE-7144
>                 URL: https://issues.apache.org/jira/browse/HIVE-7144
>             Project: Hive
>          Issue Type: Bug
>          Components: File Formats
>    Affects Versions: 0.14.0
>         Environment: ORC Table ~ 12 string columns
>            Reporter: Gopal V
>            Assignee: Gopal V
>              Labels: ORC, Performance
>         Attachments: HIVE-7144.1.patch, orc-string-write.png
>
>
> When ORC string dictionary writes data out, it suffers from bad GC performance due to a few allocations in-loop.
> !orc-string-write.png!
> The conversions are as follows
> StringTreeWriter::getStringValue() causes 2 conversions
> LazyString -> Text (LazyString::getWritableObject)
> Text -> String (LazyStringObjectInspector::getPrimitiveJavaObject)
> Then StringRedBlackTree::add() does one conversion
> String -> Text
> This causes some GC pressure with un-necessary String and byte[] array allocations.



--
This message was sent by Atlassian JIRA
(v6.2#6252)