Fundamental issues in designing data - parallel data ow computers

This paper analyses the fundamental primitives in the data-parallel computational model and proposes architectural solutions to these within the framework of a dataparallel data ow computer. The collective behaviour of this paradigm enables the use of a novel caching mechanism to be used in conjunction with an ETS matching store. It is also shown how collective behaviour may be exploited in optimising communication, through the use of a cross synchroniser, and in optimising the scheduling of reduction operations by avoiding unnecessary synchronisations. Sufcient implementation detail is given for each of these solutions to demonstrate their feasibility.