In range partitioning where partition sizes differ, data skew increases response time; some parallel execution servers take longer than others to finish their joins. Oracle recommends the use of hash partitioning and subpartitioning to enable partition-wise joins because hash partitioning, if the number of partitions is a power of two, limits the risk of skew. Ideally, the hash partitioning key is unique to minimize the risk of skew.
The number of partitions used for partition-wise joins should, if possible, be a multiple of the number of query servers. With a degree of parallelism of 16, for example, you can have 16, 32, or even 64 partitions. If there is an odd number of partitions, then some parallel execution servers are used less than others. For example, if there are 17 evenly distributed partition pairs, only one pair works on the last join, while the other pair has to wait. This is because, in the beginning of the execution, each parallel execution server works on a different partition pair. After this first phase, only one pair remains. Thus, a single parallel execution server joins this remaining pair while all other parallel execution servers are idle.
In some situations, parallel joins can cause remote I/O operations. For example, on Oracle Real Application Clusters environments running on MPP configurations, if a pair of matching partitions is not collocated on the same node, a partition-wise join requires extra internode communication due to remote I/O. This is because Oracle Database must transfer at least one partition to the node where the join is performed. In this case, it is better to explicitly redistribute the data than to use a partition-wise join.