Comments on Fundamental Issues

Seem to be easiest for PIM architectures for classic geometrically data parallel problems -- even if adaptive and irregular (these are “engineering” issues)
Must extend MPI HPF .. HPJava to support several layers of memory hierarchy
Need to study data movement in key algorithms and how they can be supported in all prototypical architectures
Can possibly use “load balancing ideas” (developed for decomposition onto CPU’s today) to address fine grain memory hierarchy decomposition in PetaFlop architectures
Need layered software model with capability to “escape” to lower levels