We describe a high-level system for granular data-parallel programming called GLU in which parallel applications are described as succinct implicitly parallel intensional compositions using sequential imperative functions. We show how different architecture-independent parallel programs can be generated from the high-level ap plication description. We also show how GLU enables functional debugging of parallel applications without. requiring their parallel execution. To illust.rate the efficiency of parallel programs generated in GLU. we consider the results of a performance study of three real parallel GLU applications, executed on two different. parallel computers. Finally, we compare GLU to other very high-level systems for granular data-parallel programming.

• concurrency95-preprint.ps