The Spar/Java Programming Language

In recent years, embedded systems with multiple processors have become increasingly important. These systems often consistof a general-purpose processor and one, oreven several, digital signal processors (DSPs). For portability, flexibility, and robustness it is often useful to regard such a cluster as a single, parallel, system. Since such a system contains several types of processors, it is said to be heterogeneous. In its simplest form, parallelization of applications can be done using an explicit parallelprogramming interface, such as provided by libraries like MPI, or Java threads. Unfortunately, programming using this model requires a detailed understanding of the complex interplay between program and machine, which is extremely laborious and error-prone. Moreover, this approach is often not portable. Instead, in theSpar/Javalanguage wefollow amore implicitapproach, whereparallelism is generated by the compiler, supported by user annotations to the program. We have designed the Spar/Java supports Java, augmented with the each and foreach constructs for semi-implicit parallelization. The each and foreach expose opportunities for parallelization to the compiler, but the details of the parallelization are left to the compiler. This allows for more robust and portable parallelization. The each and foreach statements, together with with language extensions for multi-dimensional arrays and support for more specialized arrays, form the ‘Spar’ language extensions. The Spar/Java language is described in more detail in [1, 4]. Furthermore, Spar/Java includes an annotation language [2], that allows the description of explicit or partially explicit placement of data and tasks on processors in a homogeneous or heterogeneous parallel system. In Spar/Java, all placements are static at run-time; no object migration between processors is done. The annotations do not alter the semantics of the program, they just serve as hints for the scheduler in the compiler. The annotations allow the placement of both tasks and data to be specified in a uniform manner. This research is supported by NWO (project ’Automap’), Esprit (LTR Project ‘Joses’ (#28198)), and Delft universityoftechnology(DIOCproject ‘Ubicom’).