Developing dataflow algorithms
- Los Alamos National Lab., NM (United States)
- Colorado State Univ., Fort Collins, CO (United States). Dept. of Computer Science
Our goal is to study the performance of a collection of numerical algorithms written in Id which is available to users of Motorola's dataflow machine Monsoon. We will study the dataflow performance of these implementations first under the parallel profiling simulator Id World, and second in comparison with actual dataflow execution on the Motorola Monsoon. This approach will allow us to follow the computational and structural details of the parallel algorithms as implemented on dataflow systems. When running our programs on the Id World simulator we will examine the behaviour of algorithms at dataflow graph level, where each instruction takes one timestep and data becomes available at the next. This implies that important machine level phenomena such as the effect that global communication time may have on the computation are not addressed. These phenomena will be addressed when we run our programs on the Monsoon hardware. Potential ramifications for compilation techniques, functional programming style, and program efficiency are significant to this study. In a later stage of our research we will compare the efficiency of Id programs to programs written in other languages. This comparison will be of a rather qualitative nature as there are too many degrees of freedom in a language implementation for a quantitative comparison to be of interest. We begin our study by examining one routine that exhibit different computational characteristics. This routine and its corresponding characteristics is Fast Fourier Transforms; computational parallelism and data dependences between the butterfly shuffles.
- Research Organization:
- Los Alamos National Lab., NM (United States)
- Sponsoring Organization:
- DOE; USDOE, Washington, DC (United States)
- DOE Contract Number:
- W-7405-ENG-36
- OSTI ID:
- 5498817
- Report Number(s):
- LA-UR-91-2440; CONF-910771--3; ON: DE91016300
- Country of Publication:
- United States
- Language:
- English
Similar Records
On the abstracted dataflow complexity of Fast Fourier Transforms
The dataflow time and space complexity of FFTs