When Sparse Applications Meet Architectures
Many scientific applications involve sparse computations on seemingly unstructured data. Such sparse formulations have inherent advantages in scaling of computational and storage costs but they often do not map well to the underlying hardware. As we approach the petascale and beyond with ensembles of many core chip multiprocessors, a key challenge concerns sustainable power-aware performance scaling of such applications, Many opportunities for addressing this challenge lie where applications meet architectures, demanding their adaptive co-evolution. We will discuss this in the context of sparsity, structure, parallelism and power.
Padma Raghavan, Pennsylvania State University