A decoupled data-driven architecture with vectors and macro actors

Abstract

This paper presents the implementation of scientific programs on a decoupled data-driven architecture with vectors and macro actors. This hybrid multiprocessor combines the dynamic data-flow principles of execution with the control-flow of the von Neumann model of execution. The two major ideas utilized by the decoupled model are: vector and macro actors with variable resolution, and asynchronous execution of graph and computation operations. The compiler generates graphs with various-sized actors in order to match the characteristics of the computation. For instance, vector actors are proposed for many aspects of scientific computing while lower resolution (complier-generated collection of scalar actors) or higher resolution (scalar actors) is used for unvectorizable programs. A block-scheduling technique for extracting more parallelism from sequential constructs is incorporated in the decoupled architecture. In addition a graph-level priority-scheduling mechanism is implemented that improves resource utilization and yields higher performance. A graph unit executes all graph operations and a computation unit executes all computation operations. The independence of the two main units of the machine allows the efficient pipelined execution of macro actors with diverse granularity characteristics.