An approach to extracting iteration-level parallelism from dataflow programs is presented. The method exploits the single-assignment principle, which guarantees that any data value has exactly one producer. To minimize interprocessor communication, the code is modified so that, at run time, each producer executes on the processor that holds the corresponding data. Overhead resulting from possibly remote read accesses is alleviated by a software technique similar to caching. The performance of this process-oriented dataflow system (PODS) is demonstrated using the hydrodynamics simulation benchmark called SIMPLE, in which a 19-fold speedup on a 32-processor architecture has been achieved.>

The Process-Oriented Dataflow System (PODS) is an execution model that combines the von Neumann and dataflow models of computation to gain the benefits of each. Central to PODS is the concept of array distribution and its effects on partitioning and mapping processes. The authors present and discuss the results of executing the classic matrix multiply algorithm (with 1024 data points) on a PODS simulator. The key result is that PODS can take advantage of the parallelism in matrix multiply using a simple automated partitioning scheme.>

Loosely-coupled MIMD architectures do not suffer from memory contention; hence large numbers of processors may be utilized. The main problem, however, is how to partition data and programs in order to exploit the available parallelism. In this paper we show that efficient schemes for automatic data/program partitioning and synchronization may be employed if single assignment is used. Using simulations of program loops common to scientific computations (the Livermore Loops), we demonstrate that only a small fraction of data accesses are remote and thus the degradation in network performance due to multi-processing is minimal.