On the parallel architecture with and , primary underflow partitioning produced the best performances in almost every situation. Exceptions were high values of (see figures 10.24 and 10.25) where primary minimum-overlaps partitioning performed better. This does not mean that the primary underflow strategy is the strategy to choose in every case but it does allow to conclude that well balanced partial join computations - this is the goal of the primary underflow strategy - are much more important for the join performance than a reduced number of overlapping intervals - the goal for which the primary minimum-overlaps strategy aims. The uniform partitioning strategies could not match the performances of the primary underflow and minimum-overlaps strategies. The costs were at times twice as high in comparison to primary underflow and minimum-overlaps partitioning (see tables 10.4 and 10.6, for example). This underlines our initial assumption that a naive partitioning approaches end in poor performances and that more sophisticated strategies, such as the underflow and minimum-overlaps strategies, are required.
The experiments of section 10.4 showed that, for a
temporal intersection join 
, XR and XQ values with
generally deliver the best or at least near-best performance results for .
A further, very significant conclusion is that condensation is not as harmful as we previously expected. In some cases it even improved the performances, especially for the primary minimum-overlaps strategies (see figure 10.49). A condensation factor a between 10 and 20 is feasible. We remember that this means that IP-tables can be reduced to between -th to -th of their original sizes. This not only accelerates the optimisation process (see figure 10.50) but also has many positive effects with respect to storage and maintenance of IP-tables.
The impact of black-out preprocessing largely depends on the profiles of the relations that participate in the join. On the parallel architecture, performances could improve up to 4% but also decrease up to 9% depending on the actual situation. The experiments in section 10.9 are certainly not sufficient to provide clear guidelines when and when not to use black-out preprocessing. However, the 4% increase and 9% decrease in performance suggest that the margins for improvements are only minor.