Memory Performance of Dataflow Programs

Abstract

This chapter reports on the experience of writing numeric, scientific algorithms in a mostly functional style. One-dimensional FFT, three-dimensional FFT, Eigensolvers, and a Monte Carlo Photon Transport code are studied. While the functional paradigm is effective in the expression of these problems, their memory space use is poor as compared with imperative implementations. The focus is on the programming language Id and its Monsoon Dataflow Machine implementation, and the Id codes are sometimes compared to Haskell or SISAL versions. Each algorithm is informally described, and its performance is treated in a qualitative fashion. These studies show that there continue to be serious problems in the performance of functional languages, especially with respect to space use. (For work on profiling space use in functional language implementations, see Chapter 10. The designers of Id have recognised this problem and have proposed non-functional extensions to enhance the efficiency and expressiveness of the language. Such extensions do improve performance, but they exact a programming cost due to the very low-level, explicit parallel programming that is required.