[HBASE-10592] Refactor PerformanceEvaluation tool

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 0.96.2, 0.98.1, 0.99.0
• Fix Version/s: 0.96.2, 0.98.1, 0.99.0, hbase-10070
• Component/s: test
• Hadoop Flags: Reviewed
• Assignee: Nick Dimiduk
• Reporter: Nick Dimiduk
• Created: 22/Feb/14 22:02
• Updated: 21/Feb/15 23:29
• Resolved: 26/Feb/14 22:02

Description

PerfEval is kind of a mess. It's painful to add new features because the test options are itemized and passed as parameters to internal methods. Serialization is hand-rolled and tedious. Ensuring support for mapreduce mode is a chore because of it.

This patch refactors the tool. Options are now passed around to methods and such as a POJO instead of one-by-one. Get rid of accessors that don't help anyone. On the mapreduce side, serialization is now handled using json (jackson is a dependency anyway) instead of the hand-rolled regex we used before. Also do away with custom InputSplit and FileFormat, instead using Text and NLineInputFormat. On the local mode side, combine 1 client and N clients into the same implementation. That implementation now uses an ExecutorService, so we can later decouple number of client workers from number of client tasks. Finally, drop a bunch of confusing local state, instead use the new TestOptions POJO as a parameter to static methods.

Issue Links

• is depended upon by: HBASE-10791 Add integration test to demonstrate performance improvement
• is depended upon by: HBASE-9953 PerformanceEvaluation: Decouple data size from client concurrency