On optimizing scalar self-rebalancing trees

Balanced trees are pervasive and very often found in databases or other systems which are built around querying non-static data. In this paper, we show that trees implemented as a collection of pointers shows bad data locality, poor cache performance and suffer from a lack of parallelism opportunities. We propose an alternative implementation based on arrays. Both implementations appear to be equivalently efficient time-wise. This new layout exposes new parallelism opportunities which can be then exploited by an optimizing polyhedral compiler.