Dataflow concepts have recently re-emerged in computational science as algorithm designers attempt to make programming of Chip Multiprocessors (CMPs) more tractable. The new generation of dataflow has been relabeled 'DAG scheduling' and even 'actor based programming models', but the underlying principle is the same -- by isolating side- effects, and enforcing constraints based on data-dependencies, you can efficiently and portably schedule parallel work on chip multiprocessors. A new generation of programmers appear to be re-discovering dataflow techniques as an elegant approach to managing parallelism. Moreover, the new generation is realizing these benefits using mainstream programming languages on conventional CMPs -- no need for exotic computer architectures or languages. This Minisymposium studies the evolution dataflow programming concepts from the hotbed of research activities that occured during its apex in the 1980's to modern applications.
Organizer:
John Shalf
Lawrence Berkeley National Laboratory / NERSC
Parry Husbands
Interactive Supercomputing