The PRISMA Experiment

The design of a large database management system on top of an object-oriented platform which is implemented on a tightly coupled multiprocessor system requires a large number of design decisions. There exist no methodology to evaluate the consequences and compare the differences between them. Looking into experimental methodology in physics we find a lot of similarities and techniques we think that are useful to solve our problems. A discussion is presented in which way such experiments have to be set up and of the tools that are required. The status of the project and our experience up till now is presented. 1 I n t r o d u c t i o n In the last five years the possibility to build parallel computers has become more and more appealing. On the one hand, switching speed of electronic components has increased to such a height that the speed of light poses severe limitations on the physical length of interconnections in a computer. On the other hand, integration technology of semiconductors will soon reach a point where a complex conventional computer architecture only occupies a fraction of the effective space on a silicon chip. One of the possibilities to fill up the available space on a silicon chip is to design regular parallel computers~ where a basic processor and communication design can be copied as often as required. Although the hardware possibilities are there to build a parallel computer system, the models and the experience to effectively exploit parallelism in computer systems is still to a great extent missing. The Japanese Fifth Generation Computer System Project [1] had a large influence convincing the computer science community that parallel systems had a great potential. This project was aiming at exploiting implicit~ more fine-grained parallelism, as was believed to be present in a language as Prolog. The ESPRIT project 415 [2], headed by Philips Research Laboratory in Eindhoven, also had as its target designing a parallel computer, however, its aim was to exploit explicit, more coarse-grain parallelism as present in application programs. For that purpose a language called Parallel ObjectOriented Language [3] was designed. With the help of this language the user can structure his application programs as a collection of processes (objects), which can only interact explicitly via the exchange of messages. The combination of parallel machine and language provided an object-oriented platform called Distributed Object-Oriented Machine (DOOM) [4]. It was felt that to gain more experience with explicit parallelism, it was necessary to construct large applications. Therefore the PRISMA project [5] was started as a collaboration between Philips Research Laboratory in Eindhoven and a number of Computer Science Departments at Dutch universities and the CWI, the Centre for Mathematics and Computer Science. Although targetting to design a Database and Knowledgebase machine, it became dear in the course of the project that only parallelism in databases was understood suffidently well to be able to design a parallel implementation. This paper describes the evolution in our approach and way of thinking about parallelism during the course of the four year PRISMA project. More specifically, attention wilt be paid to how we think The PRISMA project is supported by the Dutch "Stimuleringsprojectteam Informaticaonderzoek" (SPIN)