High-Productivity Stream Programming For High-Performance Systems

Applications that are structured around some notion of a " stream " are increasingly prevalent to common computing practices, and there is evidence that streaming media applications already consume a substantial fraction of the computation cycles on consumer machines [6]. Furthermore, stream processing—of voice and video data—is central to a plethora of embedded systems, including hand-held computers, cell phones, and DSPs. The stream abstraction is also fundamental to high-performance systems such as intelligent software routers, cell phone base stations, and HDTV editing consoles. Despite the prevalence of these applications, there is surprisingly little language and compiler support for practical, large-scale stream programming. Of course, the notion of a stream as a programming abstraction was established decades ago [1], and a number of special-purpose stream languages exist today (see [8] for a review). Many of these languages and representations are elegant and theoretically sound, but they are often too inflexible to support straightforward development of modern stream applications, or their implementations are too inefficient to use in practice. Consequently, most programmers resort to general-purpose languages such as C or C++ to implement stream programs. Yet there are several reasons why general-purpose languages are inadequate for stream programming. Most notably, they do not provide a natural or intuitive representation of streams, thereby reducing readability, robustness, and programmer productivity. Moreover, because the widespread parallelism and regular communication patterns of data streams are left implicit in general-purpose languages, compilers are not stream-conscious and do not perform stream-specific optimizations. As a result, performance-critical loops are often hand-coded in a low-level assembly language and must be re-implemented for each target architecture. This practice is labor-intensive, error-prone, and very costly. The StreamIt language and compiler effort at MIT is geared toward boosting programmer-productivity while con-comitantly delivering high-performance for a wide array of computing targets. The StreamIt language features several novelties that are essential for large scale program development: the language is modular, parameterizable, malleable and architecture independent. In addition, the language exposes the widespread parallelism and communication patterns that are inherent in many streaming programs, and as a result, the compiler can apply aggressive optimizations that are infeasible to perform when using conventional languages. The StreamIt compiler can deliver high performance codes—with speedups ranging up to 4.5×—for a wide array of computing targets, including embedded and desktop uniprocessors (e.g., StrongARM, IA32, and IA64), tiled and multicore architectures (e.g., Raw [11]), or grid computing systems (e.g., a cluster …