NormalizedKeySorter's and FixedLengthRecordSorter's swap and compare methods use divisions (which take a lot of time ) to calculate the index of the MemorySegment and the offset inside the segment. Greg Hogan reported on the mailing list  measuring a ~12-14% performance effect in one case.
A possibility to improve the situation is the following:
The way that QuickSort mostly uses these compare and swap methods is that it maintains two indices, and uses them to call compare and swap. The key observation is that these indices are mostly stepped by one, and incrementally calculating the quotient and modulo is actually easy when the index changes only by one: increment/decrement the modulo, and check whether the modulo has reached 0 or the divisor, and if it did, then wrap-around the modulo and increment/decrement the quotient.
To implement this, InMemorySorter would have to expose an iterator that would have the divisor and the current modulo and quotient as state, and have increment/decrement methods. Compare and swap could then have overloads that take these iterators as arguments.