Next: Partitioning Strategies
Up: Evaluation of Characteristics
Previous: Experiments
We summarise the main conclusions that can be drawn form the results
that have been described in the previous section:
- 1.
- In all experiments clearly dominates ; can be ignored. This has two consequences:
- It proves that the performance model is largely independent of
the underlying parallel programming paradigm, such as message-passing
or multithreading, as this fact is mainly relevant for the
repartitioning stage.
- When designing partitioning strategies, we can concentrate on
optimising .
- 2.
- The CPU time dominates the costs for the three subjoins.
Memory costs can become important in the case of a hardware platform
in which a large number of processors shares the access to the common
main memory (see experiment 1).
- 3.
- The subjoins that involve replicated tuples, i.e. subjoins (a)
and (c), become increasingly important
- for large numbers, m, of fragments (see experiment 2),
- for large average interval lengths, (see experiment 4).
We have to keep these major influences in mind when designing
partitioning strategies for non-uniform situations which have to
be expected in reality.
Next: Partitioning Strategies
Up: Evaluation of Characteristics
Previous: Experiments
Thomas Zurek