Because changes to the database (DB) and workload occur during a DB system's lifetime, the physical DB design must evolve to sustain good performance. These changes are carried out by on-line reorganizations which access the DB and execute concurrently with the DB workload. Dierent performance intrusions are placed on the workload when a reorganization is assigned dierent priorities compared to the workload processes. Our work studies the eects of the reorganization priority-level on performance. There are two performance aspects of interest: the time taken for a reorganization and the intrusion the reorganization places on the workload. The reorganization we consider is a redistribution (rebalancing) of a relation across the nodes of a parallel sharednothing system. A shared-nothing system contains multiple processors with their own local memory and disks, and these processors share only the interconnection network. We have seen that heavily loaded systems or highly imbalanced r...
Daniel C. Zilio