[HADOOP-12012] Investigate JNI for improving byte array comparison performance

• Type: Sub-task
• Status: Open
• Priority: Minor
• Resolution: Unresolved
• Affects Version/s: 2.7.0
• Component/s: benchmarks, io, performance

Environment

All

• Assignee: Alan Burlison
• Reporter: Alan Burlison
• Created: 21/May/15 10:27
• Updated: 21/May/15 10:27

Description

HADOOP-7761 added functionality to compare byte arrays by treating them as arrays of 64-bit longs for performance. However HADOOP-11466 reverted this change for the SPARC architecture as it causes misaligned traps which causes performance to be worse rather than better.

Most platforms have a highly-optimised memcmp() libc function that uses processor-specific functionality to perform byte array comparison as quickly as is possible for the platform.

We have done some preliminary benchmarking on Solaris that suggests that, for reasonably-sized byte arrays, JNI code using memcmp() outperforms both of the current Java byte-size and long-size implementations on both SPARC and x86 . We are confirming the results and will repeat the same benchmark on Linux and report the results here for discussion.