Improve performance of VariableWidthBlock#copyPositions

[raunaqmorarka]

Instead of copying every position one at a time, copy
contiguous ranges of Slice to reduce the number of times
SliceOutput#writeBytes is called when there are sequences
in the positions being copied.

Reduce branches for populating null values

BenchmarkVariableWidthBlock.copyPositions

(nullsAllowed) (selectedPositions) (selectedPositionsCount) us/op Before us/op After
    FALSE          GROUPED             200                      7.108        4.222
    FALSE          GROUPED             1000                     48.71        22.427
    FALSE          GROUPED             8000                     476.058      186.539
    FALSE          SEQUENCE            200                      6.912        4.107
    FALSE          SEQUENCE            1000                     47.931       21.282
    FALSE          SEQUENCE            8000                     473.199      174.198
    FALSE          RANDOM              200                      8.175        7.505
    FALSE          RANDOM              1000                     53.125       51.679
    FALSE          RANDOM              8000                     519.651      478.773
    TRUE           GROUPED             200                      5.897        4.099
    TRUE           GROUPED             1000                     38.14        19.807
    TRUE           GROUPED             8000                     394.617      158.567
    TRUE           SEQUENCE            200                      6.06         3.508
    TRUE           SEQUENCE            1000                     39.77        17.536
    TRUE           SEQUENCE            8000                     413.274      149.79
    TRUE           RANDOM              200                      7.076        7.201
    TRUE           RANDOM              1000                     42.756       43.626
    TRUE           RANDOM              8000                     416.765      385.222

[skrzypo987]

I am afraid there will be a regression for certain position arrays. Is there any benchmark?

IMHO it would be best to apply some heuristics to guess if the "new" way of copying will be faster e.g. some function of fill ratio and average position length. We can also fallback to old code if the first loop uncovers highly fragmented positions array.

[raunaqmorarka]

I am afraid there will be a regression for certain position arrays. Is there any benchmark?

IMHO it would be best to apply some heuristics to guess if the "new" way of copying will be faster e.g. some function of fill ratio and average position length. We can also fallback to old code if the first loop uncovers highly fragmented positions array.

I've tweaked the code a bit and posted JMH results with different selection patterns. It's doing better in all the benchmarked scenarios.

[sopel39]

I don't think offsets are guaranteed to make sense for null positions (add a test?)

[raunaqmorarka]

The bytes at a particular offset in slice for a null position can be junk, we don't care about that. However, the offset positions need to be valid in existing code as well, otherwise the result of getSliceLength for the next non-null position could go wrong.

[sopel39]

lgtm % comments

[sopel39]

Why 0 is bad?

[raunaqmorarka]

Not sure, I just copied this one from BenchmarkDataGenerator#randomNullChance
@skrzypo987 do you know why ?

[skrzypo987]

There is a < at the end, so if random picks 0 it will return null for null chance of 0.
We basically want to pick random out of [0,1>, but nextBoolean returns <0,1>

[sopel39]

I would simplify it

[sopel39]

I would simplify it