It is widely claimed that functional languages are particularly suitable for programming parallel computers. A claimed advantage is that the programmer is not burdened with details of task creation, placement, scheduling, and synchronisation, these decisions being taken by the system instead. Leaving aside the question of whether a pure functional language is expressive enough to encompass all the parallel algorithms we might wish to program, there remains the question of how effectively the compiler and run-time system map the program onto a real parallel system, a task usually carried out mostly by the programmer. This is the question we address in this paper.
We first introduce the system architecture of GRIP, a shared-memory parallel machine supporting an implementation of the functional language Haskell. GRIP executes functional programs in parallel using compiled supercombinator graph reduction, a form of declarative rule system.
We then describe several strategies for run-time resource control which we have tried, presenting comprehensive measurements of their effectiveness. We are particularly concerned with strategies controlling task creation, in order to improve task granularity and minimise communication overheads. This is, so far as we know, one of the first attempts to make a systematic study of task-control strategies in a high-performance parallel functional-language system. GRIP's high absolute performance render these results credible for real applications.
@InProceedings{grip-sched,
author = {Kevin Hammond and Simon L. {Peyton Jones}},
title = {Profiling Scheduling Strategies on the {GRIP} Multiprocessor},
booktitle = {International. Workshop on the Parallel Implementation of Functional Languages},
year = {1992},
address = {RWTH Aachen, Germany},
month = sep,
pages = {73--98}
}