1. In order to reduce contention for disk access, the database can be partitioned over several disks, allowing several disk accesses to be serviced in parallel.

2. In order to exploit the potential for parallel disk access, we must choose a good distribution of data among the disks.

3. For the parallel 2-way join, it is useful to distribute tuples of individual relations among several disks (disk stripping). For example, assign tuples to disks based on the hash function value of the hash-join algorithm. All groups of tuples that share a bucket are assigned to the same disk. Each group is assigned to the same disk, if possible, or the groups are distributed uniformly among the available disks. This allows the parallel 2-way hash-join to exploit parallel disk access.

4. For the pipeline-join, it is desirable that each relation be kept on one disk and the distinct relations be assigned to separate disks to thedegree possible.

For example, for computing (r1 ∝ r2) 1 (r3 ∝ r4), if each relation is on a di erent disk, contention is eliminated between processors P1 and P2.

5. The optimal physical organization differs for different queries. The DBA must choose a physical organization that is believed to be good for the expected mix of database queries.

6. The query optimizer must choose from the various parallel and sequential techniques by estimating the cost of each technique on the given physical organization.