On the use of data-flow techniques in database machines
Haran Boral · Minds at UW (University of Wisconsin) · 1981
The past decade has seen a number of design efforts in the area of database machines. Research has shown that all of the major designs suffer from some flaw leading to the inefficient execution of one or more operations. In this thesis we show that the lack of systematic study of the algorithms to be used by an architecture before a hardware design is picked is the reason for these flawed designs. We then consider a number of possible algorithms for all the relational algebra operators and introduce a new design based a group of these. The proposed machine utilizes a local network communication mechanism and employs a data-flow strategy for query processing. Previous research has shown both advantages and disadvantages of using a data-flow query processing strategy. In particular, it was shown that data movement between the mass storage devices and processors is minimized at the expense of additional control messages. In this design we show how such a strategy can be employed without the large control overhead. We also consider the problem of associating logic with a disk for the implementation of certain operations on the fly. Three design approaches are examined and compared. It is shown how an associative disk can be incorporated into a database that supports both on-the-disk-and off-the-disk processing.