[
https://issues.apache.org/jira/browse/HIVE-15987?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Teddy Choi updated HIVE-15987:
------------------------------
Description:
Most of data operations in Hive uses null operations. The current
implementation of ColumnVector.isNull uses a boolean array, which uses 8 bits
per 1 boolean. BitSet is a more compact representation, as it uses 1 bit per 1
boolean with a backing long array. Also logical operations between longs are
much faster than ones with bytes as it uses less instructions per byte. So it
will bring 8x or more performance for null operations.
However, there also are several cases that will make this improvement slow.
Such as simple reads will require more instructions per row. So it should
include benchmark tests to show its performance impact.
was:Most of data operations in Hive uses null operations. The current
implementation of ColumnVector.isNull uses a boolean array, which uses 8 bits
per 1 boolean. BitSet is a more compact representation, as it uses 1 bit per 1
boolean with a backing long array. Also logical operations between longs are
much faster than ones with bytes as it uses less instructions per byte. So it
will bring 8x or more performance for null operations.
> Replace ColumnVector.isNull boolean[] impl. with BitSet
> -------------------------------------------------------
>
> Key: HIVE-15987
> URL: https://issues.apache.org/jira/browse/HIVE-15987
> Project: Hive
> Issue Type: Improvement
> Reporter: Teddy Choi
> Assignee: Teddy Choi
>
> Most of data operations in Hive uses null operations. The current
> implementation of ColumnVector.isNull uses a boolean array, which uses 8 bits
> per 1 boolean. BitSet is a more compact representation, as it uses 1 bit per
> 1 boolean with a backing long array. Also logical operations between longs
> are much faster than ones with bytes as it uses less instructions per byte.
> So it will bring 8x or more performance for null operations.
> However, there also are several cases that will make this improvement slow.
> Such as simple reads will require more instructions per row. So it should
> include benchmark tests to show its performance impact.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
