Programming Support for Cell/BE Multiprocessor

Abstract. An emerging class of architectures are accelerator-based het-erogeneous multiprocessors with software-managed memory hierarchies like Cell/BE. A major difficulty in programming such kind of machines are the explicit data transfers between the different memories raising new programming challenges. In this paper we discuss a programming approach which supports application programmers in writing efficient code for non-cache-based architectures. A crucial role plays the interplay between the three parties programmer, parallelization framework, and native compiler. Based on our experiences with past programming ap-proaches, we propose language extensions to orchestrate parallel execu-tion of threads and to control data transfers. Experiments are performed to analyze the roles of programmer and compiler in more detail.