Relational and Entity-Relationship Model Databases and Specialized Design Files in VLSI Design

Databases have a number of advantages over specialized design files for use in the VLSI design process. Balanced against these advantages is the presumed extra run-time cost of accessing the database. In particular, ordinary relational databases appear to be too slow to be used on-line by engineers in the design process [Eastman80] [Haynie81] [Sidle80] [Stonebraker82]. The requirements for design access include the retrieval of major units of data at one time, implying joins of many attributes with their libraries, and projections to obtain the relevant attributes for a given study. Previously, in order to test if databases could achieve acceptable performance at all, we experimented using a CODASYL database system [Beetem82] [Wiederhold82]. This appeared to provide adequate performance, but at a high cost of software maintenance and lack of flexibility due to CODASYL limitations. Our hypothesis is that the relational data model per se is not the crucial factor in performance; rather, the internal access mechanisms of the database significantly affect performance. To test this hypothesis, we ran experiments using the Cypress database system in the Cedar environment at Xerox Palo Alto Research Center. The program used is a design macro expander, more fully described in Section 3. The tests show that colocating pointers to all relations that a VLSI component appears in will reduce CPU time spent in the main database access routines by 25%. We repeated a subset of these experiments on a Digital VAX 750 using the Structured Design System (SDS) of Silvar-Lisco [SL83] with specialized design files rather than a database management system, and found that when a structured result file is used, the largest component of execution time is the time spent writing out the expansion results.