The Assimilation of Dataflow
				   
			   David E. Culler
		      Computer Science Division
		  University of California, Berkeley
				   
				   
Dataflow models of computation, dataflow machines, and dataflow
languages have been actively studied and developed over the past
fifteen years.  This work has been approached as a ``new paradigm'' of
computing, i.e, self-consistent and powerful, although the novel
machines might be ill-suited to conventional languages and the novel
languages ill-suited to conventional machines.  The work seemed to be
driven by a unique set of critical issues.

This talk will discuss how the dataflow paradigm has assimilated into
the mainstream of parallel computing and examine the common set of
critical issues.  First, we will look at how the dataflow language
Id90 can be implemented efficiently on conventional distributed memory
multiprocessors and address some fundamental shortcomings of dataflow
machines in the process.  This work involves sophisticated compilation
techniques that operate on dataflow graphs to produce a collection of
self-scheduling threads.  Second, we will look at how latency
tolerance can be incorporated into conventional parallel languages.
Finally, we will describe a simple communication primitive for
parallel machines, called active messages. This provides a basis for
these language implementations and more conventional approaches, such
as message passing or shared memory.