An abstract machine for parallel graph reduction

An abstract machine suitable for parallel graph reduction on a shared memory multiprocessor is described. Parallel programming is plagued with subtle race conditions resulting in deadlock or fatal system errors. Due to the nondeterministic nature of program execution the utilization of resources may vary from one run to another. The abstract machine has been designed for the efficient execution of normal order functional languages. The instructions proposed related to parallel activity are sensitive to load conditions and the current utilization of resources on the machine. The novel aspect of the architecture is the very simple set of instructions needed to control the complexities of parallel execution. This is an important step towards building a compiler for multiprocessor machines and to further language research in this area. Sample test programs hand coded in this instruction set show good performance on our 18 node BBN Butterfly as compared to a VAX 8600.