Dataflow is one of the major models of parallel computation. Implementation of a General Purpose Dataflow Multiprocessor extends work in this area by introducing an unusually simple model of dynamic dataflow execution, called the Explicit Token Store (ETS) architecture, and its realization in Monsoon, a large-scale dataflow multiprocessor. Monsoon is currently under construction at the Motorola Microcomputer Division. Papadopoulos argues that the underlying sequential architecture of contemporary multiprocessors has not been able to support the synchronization demands of parallel execution and that these systems have largely failed to meet expectations for programmability and performance. He points out that processors must be fundamentally changed to execute a parallel machine language that coordinates parallel activities efficiently as instructions are scheduled. Although dataflow architectures have met this challenge by radically reformulating the basic specification of a machine program, they have suffered from substantial implementation shortcomings, notable the need for large associative memories. The ETS architecture Papadopoulos introduces here achieves the power of previous tagged-token dataflow architectures, but with a much leaner cycle and much less complexity. Gregory Papadopoulos is an Assistant Professor of Electrical Engineering and Computer Science in the Laboratory for Computer Science at MIT. Contents: General Purpose Multiprocessing. The TaggedToken Dataflow Architecture. The Explicit Token Store. Compiling for an ETS Dataflow Processor. Compiling Imperative Languages for an ETS. Monsoon: An ETS Multiprocessor. A Monsoon Instruction Decoding.
General purpose multiprocessors have largely failed to meet expectations for programmability and performance. The author blames the lack of usable parallel programming languages and systems on the underlying processor architecture. Machines built out of conventional sequential processors simply do not support the synchronization demands of parallel execution so the programmer focuses upon the dangerous and arduous task of discovering a minimum set of sychronization points without introducing nondeterminisim. He argues that processors must be fundamentally changed to execute a parallel machine language, in which parallel activities are coordinated as efficiently as instructions are scheduled. Dataflow architectures address this challenge by radically reformulating the basic specification of a machine program. These machines directly execute dataflow graphs, which specify only the essential prerequisites for the execution of an instruction - the availability of operands.
Unfortunately, dataflow machines, including the MIT Tagged Token Dataflow Architecture (TTDA), have laboured under a number of implementation burdens, notably the apparent need for a fully associative operand matching store which discovers when instructions are able to execute. This work introduces and develops a novel dataflow architecture, the Explicit Token Store (ETS), which directly executes tagged-token dataflow graphs while correcting a number of inherent inefficiencies of previous dataflow machines. In the ETS model, operand matching is performed at compiler-designated offsets within an activation frame. It shows that the ETS is compatible with the TTDA by giving translations from TTDA machine graphs to ETS machine graphs. Finally, this work describes an implementation of an ETS dataflow multiprocessor, called Monsoon, now under construction.
