1 a Framework for Implementing Highlyparallel Applications on Distributedmemory

We present here a massively parallel formalism, the Cellular Abstract Machine (CAM), and its implementation principles on a distributed memory multi-processor architecture. The theoretical machine is more especially dedicated to connection-nist applications and other massively parallel artiicial intelligence applications, such as semantic networks. This paper brieey describes the CAM and gives an example of CAM program. Then we explain an implementation method on transputer based architectures. We derive theoretical expressions predicting the implementation performances. We nally discuss brieey some practical issues of the actual multi-processor implementation. The concept of an abstract or virtual machine has been widely used since the beginning of modern computer science history. Its rst advantage is to hide from the developer the details of the real computer's architecture. The second advantage of the abstract machine concept is to increase software portability to a new computer architecture. Only the virtual machine itself has to be actually ported and not the languages and the applications running in it. When it comes to parallel computing, the speciic features of the hardware innuence software development even more than in the case of sequential computing. Placement of the concurrent parts of the software on the diierent hardware components, load balancing and communication optimisation become major issues of the development. They sometimes induce programmers to re-parallelize parallel applications in a way diier-ent from their initial form. Eventually, in the course of solving a parallel