[
https://issues.apache.org/jira/browse/ORC-574?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
David Zanter updated ORC-574:
-----------------------------
Summary: Performance: Use const references for string statistics min and
max to avoid copy construction (was: Performance: Statistics getMax/getMin
should return const vals to prevent std:string copies)
> Performance: Use const references for string statistics min and max to avoid
> copy construction
> ----------------------------------------------------------------------------------------------
>
> Key: ORC-574
> URL: https://issues.apache.org/jira/browse/ORC-574
> Project: ORC
> Issue Type: Improvement
> Components: C++
> Affects Versions: 1.6.2
> Reporter: David Zanter
> Priority: Major
> Attachments: callgrind-before-after.JPG
>
>
> Via Callgrind Performance Profiling of a scenario of a Copy (Full Read and
> then Full Write) of a 1.9 million row ZLib Compressed ORC Table. The #4
> Usage of CPU is the std::string alloc from being called by:
> orc::StringColumnStatisticsImpl::update method due to the getMax/getMin calls
> causing std:string alloc/copy/delete.
>
> Changing the getMaximum/getMinimum methods to return const vals will prevent
> these alloc/copy/deletes from occurring.
>
> Currently with 1.6.X master the performance profile of this scenario is:
> Instructions Executed: 16.6 Billion Instructions
> real clock time 3.91 seconds
>
> With the fix to use consts, this improves the CPU usage by about 38% and the
> Clock Time about 10% to:
> Instructions Executed: 12.0 Billion Instructions
> real clock time 3.53 seconds
>
> Attached JPG showing before (left) and after (right) screenshot of callgrind.
>
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
