Schedule

Keynote by Barbara Chapman: An Evolutionary Exascale Programming Model Deserves Revolutionary Support

An evolutionary approach to large-scale application development would simplify application migration and help preserve investment in legacy code where feasible. Many of today's DOE MPI-based application codes employ OpenMP directives to reduce the memory footprint within system nodes, and to improve the communication/computation ratio of the MPI processes. Yet OpenMP has a number of weaknesses that make the current version insufficient for the expression of parallelism to exploit emerging and future node architectures. In this presentation we discuss possible means to enhance this shared memory programming interface to enable its use with MPI or another inter-node programming API on larger, more complex high-end computer architectures.

If such a model is to provide high performance along with productivity benefits, our efforts must go beyond the language level. We consider what innovation at the compiler level and in the runtime system might improve the mapping of a code to a specific hardware and execution environment, and continue to adapt it to changing circumstances, without the need for manual control or external intervention.