Distributed Types: A unifying model of spatial structure in parallel programming

Andreas Schramm

In programming massively-parallel machines, preserving locality is the major issue (because of the physical communication incurred).

Thus, the intent of Distributed Types is to formalize the spatial structure of data-parallel applications explicitly at a problem-specific level.

The advantages of expressiveness at a high-level are that the low-level implementation aspects such as mapping and message generation can be delegated to the system. The benefits that can be expected are more compact and comprehensible programs and improved architecture independence.

The actual challenge of this approach is that the formalization should reflect the structuring principles that actually occur in the relevant class of applications.

The presented talk identified the most important problems that massively-parallel applications, mainly numeric ones, pose with respect to spatial structures, and introduced a solution.