TI A Multilayered Data Flow Computer Architecture 

AU John Gurd 

AU Ian Watson 

AU John Glauert 

LT UMCS-80-3-1 

OR UMCS 

AV http
http:www.cs.man.ac.ukcsonlycstechrepAbstractsUMCS-80-3-1.html 

AV email techreports@cs.man.ac.uk 

MN March 

YR 1980 

AB This paper introduces a multilayered data flow computer
architecture based upon a high level language and underlying
graphical notation for representing computations.    The
language, LAPSE, employs a single assignment rule and several
more familiar parallel programming constructs.  It permits both
iteration and recursion.    The graphical notation utilises the
concept of labels for those tokens which use computational
subgraphs reentrantly.  The operation of labels in implementing
iteration, recursion and data structures is illustrated. The
architecture is based on a ring-structure in which binary
representations of tokens, known as results, circulate.  A
processing unit computes results as required by a stored program
which represents the computational graph.  Each result carries
with it a name which is derived from the notational label.  Names
perform two functions : Firstly they separate different
instantiations of reentrant code;  secondly they can be
associatively matched with other inputs to their dstination node,
to determine when the node may be executed. The ring-structure
can be pipelined to achieve an instruction execution rate which
is limited by the amount of parallel activity permitted by the
program and the technology used.    An extensible version of the
architecture contains many such rings in a multilayered structure
whose execution rate is bound solely by the inherent parallelism
of the programs running on it.  The architecture can also be
organised so as to exhibit a high degree of tolerance to hardware
component failures.