In the past many applications have ensured success by restricting the size of the application, or by increasing the number of processors and memory size to enable the full database to be supported. Here, we specify that databases of arbitrary sizes should be supported and not be restricted by the memory size of individual processors.The ability to cope with very large databases was easily achieved in many of the early MIMD systems by using a shared memory model. However, the transputer and Occam process model restricts us from using this approach, instead we may share data [7].Unlike shared memory systems, we cannot globally address data in a message passing system. However, if data items carry unique identifiers, we can share single or multiple copies of those data items across many processors. Indeed, adopting this system of shared data reference allows us the same memory flexibility for read-only data, as would be obtained in a shared memory system, without the bus contention problems associated with that class of processor. In its degenerate form, a shared data system has only private data, which is never available at any other processor. The simple processor farm of May and Shepherd [8] is a typical example, where data and tasks are assigned to specific processors without the need for data to migrate to other processors. In many applications, such as the ray tracing of very complex computer images, a static allocation of data is inappropriate. Here, a database is managed at each node in a similare manner to a cache memory. Shared data systems for a tree based system architecture, and for very large data base problems are described by Green, Paddon and Lewis [7], and Green and Paddon [3, 4, 5, 6], where these systems were applied to image synthesis using the ray tracing method.