[MAPREDUCE-2629] Class loading quirk prevents inner class method compilation

• Type: Improvement
• Status: Closed
• Priority: Minor
• Resolution: Fixed
• Affects Version/s: 0.21.0, 0.22.0
• Fix Version/s: 0.23.0
• Component/s: task
• Target Version/s: 0.23.0
• Hadoop Flags: Reviewed
• Assignee: Eric Caspole
• Reporter: Eric Caspole
• Created: 28/Jun/11 22:06
• Updated: 15/Nov/11 00:48
• Resolved: 21/Oct/11 05:18

Description

While profiling jobs like terasort and gridmix, I noticed that a
method "org.apache.hadoop.mapreduce.task.ReduceContextImpl.access
$000" is near the top. It turns out that this is because the
ReduceContextImpl class has a member backupStore which is accessed
from an inner class ReduceContextImpl$ValueIterator. Due to the way
synthetic accessor methods work, every access of backupStore results
in a call to access$000 to the outer class. For some portion of the
run, backupStore is null and the BackupStore class has never been
loaded by the reducer.

Due to the way the Hotspot JVM inliner works, by default it will not
inline a short method where the class of of the return value object
is unloaded - if you use a debug JVM with -XX:+PrintCompilation you
will see a failure reason message like "unloaded signature classes."
This causes every call to ReduceContextImpl.access$000 to be executed
in the interpreter for the handful of bytecodes to return the null
backupStore.