TAJO-1407 Minor performance improvement of MemSortExec

[navis]

Simple micro benchmark shows 15~30% performance improvement, but this uses more memory. (Also can make little difference on appearance order of non-key values, caused by different sort implementation)

[navis]

Removed JDK7 only method

[jihoonson]

Looks interesting.
I'll review tonight.

[jihoonson]

Would you mind adding some descriptions on this class?

[navis]

Added comments and now it supports interval type.

[jihoonson]

I'll review this weekend.

[jihoonson]

TajoDataTypes.Type is the enum type of PB. Each type already has a corresponding integer.
type.getNumber() returns an integer value corresponding to the given type, and Type.valueOf() returns a type corresponding to the given integer value.
How about use these functions?

[navis]

It's possible but I prefer switch-case block to be compact as possible as it can be. And most of types in PB enum cannot be supported by this.

[jihoonson]

Ok. It looks good.

[jihoonson]

@navis, this is truly impressive work.
With my laptop, the vectorized sorter is faster than Collections.sort() about 25%!
I left some trivial comments.
In addition, it will be great if you add some tests to verify sort result.

[navis]

Revised test case, which can be fail sometimes

[jihoonson]

I run the test 5 times, and it passed every time.
Would you share when the test fails?

[navis]

2M rows sometimes made GCs, causing inversion of elapsed time.

[jihoonson]

Oh, I think it's ok. The revised test doesn't contain any comparisons of elapsed times.
In addition, MemSortExec is executed only when there remains a sufficient amount of memory.
Here is my +1, and I'll commit it shortly.
Thanks for your great work!

[jihoonson]

Oops, sorry. I have to wait the CI result.
Would you put the latest patch on Jira?
I'd like to commit this patch as soon as possible, and Jenkins will be much faster than Travis CI.