Design and implementation of 8 [formula omitted]: A declarative data-parallel language

Abstract

In this article we advocate a declarative approach to data-parallelism to provide both parallelism expressiveness and efficient execution of data intensive applications. 8 1 2 , an experimental language combining features of collection and stream oriented languages in a declarative framework, is presented. A new structure, the web, allows the programmer to write programmes as mathematical expressions and to implicitly express data and control parallelism. The first part of this paper proposes a classification of the various expressions of parallelism in programming languages. We show that hybrid execution models combining both data and control parallelism are possible and necessary to get an effective speedup. We sketch the advantage of the declarative style with respect to parallelism expression (application side) and exploitation (compiler side). In the second part we describe the 8 1 2 language and the concepts of collection, stream and web. A web is a multi-dimensional object that represents the successive values of a structured set of variables. Some 8 1 2 programmes are given to show the relevance of the web data structure for simulation applications (a resolution of O.D.P.E. and a simulation in artificial life). Examples of 8 1 2 programmes, involving the dynamic creation and destruction of webs, are also given. Such programmes are necessary for simulations of growing systems. In the third part, the implementation of a compiler restricted to the static part of the language is described. We focus on the process of web equations compilation towards a virtual SIMD machine. We also present the clock calculus, the scheduling inference and the distribution of the computations among the processing elements of a parallel computer.