In profiling the Parquet reader, the variable length decoding appears to be a major bottleneck making the reader CPU bound rather than disk bound.
A yourkit profile indicates the following methods being severe bottlenecks -
NullableVarBinaryVector$Mutator.setSafe(int, int, int, int, DrillBuf)
The problem is that each of these methods does some form of bounds checking and eventually of course, the actual write to the ByteBuf is also bounds checked.
DrillBuf.chk can be disabled by a configuration setting. Disabling this does improve performance of TPCH queries. In addition, all regression, unit, and TPCH-SF100 tests pass.
I would recommend we allow users to turn this check off if there are performance critical queries.
Removing the bounds checking at every level is going to be a fair amount of work. In the meantime, it appears that a few simple changes to variable length vectors improves query performance by about 10% across the board.