Issue No.10 - October (1980 vol.29)
K.P. Gostelow , General Electric Research and Development Center
Our goal is to devise a computer comprising large numbers of cooperating processors (LSI). In doing so we reject the sequential and memory cell semantics of the von Neumann model, and instead adopt the asynchronous and functional semantics of dataflow. We briefly describe the high-level dataflow programming language Id, as well as an initial design for a dataflow machine and the results of detailed deterministic simulation experiments on a part of that machine. For example, we show that a dataflow machine can automatically unfold the nested loops of n X n matrix multiply to reduce its time complexity from 0(n3) to 0(n) so long as sufficient processors and communication capacity is available. Similarly, quicksort executes with average 0(n) time demanding 0(n) processors. Also discussed are the use of processor and communication time complexity analysis and "flow analysis," as aids in understanding the behavior of the machine.
parallel computer, Asynchronous execution, concurrency, dataflow, distributed computer, functionality, large-scale integration, locality, multiprocessor architecture
K.P. Gostelow, R.E. Thomas, "Performance of a Simulated Dataflow Computer", IEEE Transactions on Computers, vol.29, no. 10, pp. 905-919, October 1980, doi:10.1109/TC.1980.1675474