Integrated execution: A programming model for accelerators

Abstract

The Cell Broadband Engine® (Cell/B.E.) processor was designed to provide a mix of central cores for control code and accelerators optimized for data processing. A heterogeneous design allows different processor elements to be optimized for specific functions and makes each processor element more area and power efficient. Cell/B.E. processor-based systems are the most powerful and the most power-efficient systems in the world, as represented by the Top500™ and Green500 lists. This paper offers a new view of the architectural design choices that were made in consideration of software usability and application development for the Cell/B.E. processor. Specifically, we explore the concept of integrated executables that allow a single application to execute across multiple heterogeneous processor elements. Hardware and software architectures were co-optimized to allow an application executing on multiple heterogeneous cores to efficiently communicate and share data, which is key to exploiting chip multiprocessors with ever-increasing numbers of cores.