Data Placement and Workload Balancing

• SPRINT distributes the attribute lists evenly over N processors of a shared-nothing machine. Thus each processor works on only 1/N of the total data.

• For performing parallel classification, the training-set examples are distributed equally among the N processors.

• Each processor generates its own attribute-list partitions in parallel by projecting out each attribute from the training-set examples it was assigned.

• Continuous attribute lists are required to be sorted by the SPRINT algorithm. For this parallel sorting algorithm is used. Continuous attribute lists are first sorted by each processor locally and then are are repartitioned into contiguous sorted sections.