GLU is a hybrid intensional language for program ming conventional parallel computers in which parallel programs are constructed by composing sequential, granular C functions using Lucid, a purely intensional language. Parallelism implicit in the Lucid specification is exploited at runtime using eduction, an intensional model of computation.
We describe a basic eduction-based parallel implementation of G LU consisting of a single master process and multiple stateless worker processes. Using the CM-5, we show that this GLU implementation has limited performance scalability, especially when the granularity of parallelism is not sufficiently coarse. We show how the scalability can be improved by decentralizing some state from the master to the workers. We then propose a new eduction-based parallel implementation of GLU using multiple masters that is designed to be more scalably efficient than either of the single-master parallel implementations.