[HBASE-10320] Avoid ArrayList.iterator() ExplicitColumnTracker

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 0.98.0, 0.96.2, 0.99.0, 0.94.17
• Component/s: Performance
• Hadoop Flags: Reviewed
• Assignee: Lars Hofhansl
• Reporter: Lars Hofhansl
• Created: 12/Jan/14 00:42
• Updated: 26/Feb/14 04:45
• Resolved: 14/Jan/14 06:55

Description

I noticed that in a profiler (sampler) run ScanQueryMatcher.setRow(...) showed up at all.
In turns out that the expensive part is iterating over the columns in ExcplicitColumnTracker.reset(). I did some microbenchmarks and found that

private ArrayList<X> l;
...
for (int i=0; i<l.size(); i++) {
   X = l.get(i);
   ...
}

Is twice as fast as:

private ArrayList<X> l;
...
for (X : l) {
   ...
}

The indexed version asymptotically approaches the iterator version, but even at 1m entries it is still faster.
In my tight loop scans this provides for a 5% performance improvement overall when the ExcplicitColumnTracker is used.

Edit:

private X[] l;
...
for (int i=0; i<l.length; i++) {
   X = l[i];
   ...
}

Is even better. Apparently the JVM can even save the boundary check in each iteration.