A process-oriented model for efficient execution of dataflow programs

Abstract

Dividing computation into subtasks that can be executed on separate processing elements is a very difficult task. Dataflow systems represent an extreme case where each machine instruction is an independent subcomputation. As a consequence, the execution overhead is very high. In this paper, we present an execution model for dataflow where the unit of computation is not a single instruction. Rather, the dataflow graph is divided into paths according to their data dependencies. Each path is then treated as a very simple process: it is loaded into memory; it switches between ready, running, and block states; and it communicates with other such processes through messages. The main advantage of the proposed approach over conventional approaches to parallelism is that there is a mechanical way of creating subcomputations that can be executed in parallel. At the same time, this approach does not suffer from the inefficiencies inherent to purely dataflow systems. Instead, it permits the granularity of processes to be adjusted to achieve a balance between the amount of parallelism that can usefully be exploited and the amount of sequential execution that can be handled effectively within one process.