Utilizing machine topology and heterogeneity in numerical algorithms

Luke Olson, University of Illinois at Urbana-Champaign

Usage Details

Sparse matrix operations abound in numerical simulations and represent significant costs at scale. As we anticipate future machine networks and compute units, these communication-bound operations will continue to incur significant cost. The focus of this allocation proposal is on reducing communication at scale, particularly in settings where machine layout and multiple compute units can be exploited. Blue Waters is an ideal setting for developing these methods since we can expose node-level parallelism (as well as socket-level) and identify computations suitable for GPU acceleration. Specifically, we focus on three interrelated aspects: network contention in both unstructured and structured solvers, identifying node-level optimizations, and managing accelerators in this context.