Rekindling Parallelism

Computing in parallel means performing computation simultaneously, this generates two distinct views: (1) Performance view A mean to accelerate computation using coarse grain parallelism. (2) Decentralization view A new way of programming by decentralizing massive fine grain parallelism. Researchers on massive parallel models study the programming expressiveness, i.e. new bio-inspired ways of computing such as artificial neural network or multi agent systems solving new kinds of problems, but are usually not directly concerned about high performance. In contrast, researchers on high performance tend to narrow the scope of parallel expressiveness by preserving the sequential model of computation and defining specific language constructs that can lead to parallel run-time performance for more classical parallel algorithms. We argue that parallelism will really fully blossom only when both views get unified through the achievement of a new generic computing model that, while enabling decentralized computation, also supports classical way of programming and incorporates the hardware constraints to provide parallel performance. We are working on such a generic model called self developing self mapping network. This paper first justifies the motivation for such a model, and then sketches the fundamental principles of this model.