[SPARK-12950] Improve performance of BytesToBytesMap

• Type: Improvement
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 2.0.0
• Component/s: SQL
• Epic Link: Whole stage codegen
• Assignee: Davies Liu
• Reporter: Davies Liu
• Created: 21/Jan/16 05:43
• Updated: 11/Feb/16 09:44
• Resolved: 10/Feb/16 00:44

Description

When benchmark generated aggregate with grouping keys, the profiling show that lookup in BytesToBytesMap took about 90% of the CPU time, we should optimize it.

After profiling with jvisualvm, here are the things that take most of the time:

1. decode address from Long to baseObject and offset
2. calculate hash code
3. compare the bytes (equality check)

Issue Links

• links to: [Github] Pull Request #10877 (viirya)
• links to: [Github] Pull Request #11002 (viirya)
• links to: [Github] Pull Request #11010 (davies)