Citation Context

...ty is at a sufficiently high level such that the nature of the underlying architecture (shared memory versus distributed memory) is hidden. It is also possible to use systems such as Linda [4] or PVM =-=[8]-=- or Active Messages [9] to provide this support. 4 Application Kernels in GLU Having briefly described the GLU programming system, we now illustrate the expressiveness of GLU in two ways: in being abl...

Citation Context

... The facility is at a sufficiently high level such that the nature of the underlying architecture (shared memory versus distributed memory) is hidden. It is also possible to use systems such as Linda =-=[4]-=- or PVM [8] or Active Messages [9] to provide this support. 4 Application Kernels in GLU Having briefly described the GLU programming system, we now illustrate the expressiveness of GLU in two ways: i...

Citation Context

...rogram composition and a C part that defines various data types and data functions referred to in the Lucid part. 2.1 Lucid circa 1994 The language Lucid has evolved from a temporal dataflow language =-=[10]-=- to a multidimensional dataflow language [3]. It is the only dataflow language in which multidimensional data structures are implicit which has significant consequences on its ability to express diffe...

Citation Context

...macy over all possible models of computation that can be used to derive its meaning operationally. The model of computation embodied by the generator/executor abstract architecture is called eduction =-=[1]-=-. Eduction corresponds to a lazy evaluation strategy for Lucid (and GLU) programs. The strategy is typically implemented using demand-driven execution although it is also possible to use a combination...

Citation Context

...ataflow programming language Lucid by extending Lucid in two simple ways: user-defined functions are specified in a foreign language (such as C) and values are of foreign types (such as C data types) =-=[5, 3]-=-. Programming a conventional parallel computer with GLU consists of three stages: 1. Develop a coarse-grain dataflow program using GLU in which parallelism is implicitly expressed. 2. Select an approp...

Citation Context

...strategy for Lucid (and GLU) programs. The strategy is typically implemented using demand-driven execution although it is also possible to use a combination of demand-driven and data-driven execution =-=[6]-=-. Using a lazy model of computation such as eduction means that there is no superfluous computation. However, exploitation of parallelism is more conservative when compared to its eager counterpart. A...

Citation Context

... high level such that the nature of the underlying architecture (shared memory versus distributed memory) is hidden. It is also possible to use systems such as Linda [4] or PVM [8] or Active Messages =-=[9]-=- to provide this support. 4 Application Kernels in GLU Having briefly described the GLU programming system, we now illustrate the expressiveness of GLU in two ways: in being able to succinctly express...

Citation Context

...ataflow programming language Lucid by extending Lucid in two simple ways: user-defined functions are specified in a foreign language (such as C) and values are of foreign types (such as C data types) =-=[5, 3]-=-. Programming a conventional parallel computer with GLU consists of three stages: 1. Develop a coarse-grain dataflow program using GLU in which parallelism is implicitly expressed. 2. Select an approp...

Citation Context

...e considered, we have seen repeated expression of tournament parallelism using linear tree and cubic tree. This is because tournament parallelism is inherent in several otherwise diverse applications =-=[7]-=-. 5 Parallel Execution of GLU Programs The principal objective of this section on performance is to show that GLU programs despite being expressed at a very high-level of abstraction do not pay for th...

Citation Context

...and y contexts using operations prev and next. 2.2.3 Tournament Parallelism Tournament parallelism arises in tree computations as logarithmically decreasing data parallelism at each level of the tree =-=[2]. (It is c-=-losely related to parallelism found in "divide-and-conquer" algorithms.) In Lucid, tournament parallelism is an example of dimensional parallelism as illustrated by the following example. c ...