Our SC14 HPGMG birds of a feather (BoF) session took place today. Mark Adams gave an introduction to our benchmarking effort, Sam Williams gave details about the finite-volume implementation and performance results, and Jed Brown highlighted characteristics of the finite-element implementation and the concept of "dynamic range" or "performance versatility". Following these presentations by the primary developers, Mitsuhisa Sato from RIKEN AICS presented recent experience on the K computer and Muthu Baskaran of Reservoir Labs presented recent work on task-based programming models. The slides for all talks are available.
The BoF concluded with audience questions for the panel, notably
-
Q: What about nonlinear problems such as Poisson-Boltzmann? A: Nonlinearities create a scale dependence, but we want a scale-independent benchmark (where convergence does not depend on resolution) and do not believe that a weak nonlinearity changes the machine characterization. See further discussion.
-
Q: Which MPI primitives are used? A: Both implementations use nonblocking point-to-point by default. HPGMG-FE has an option to use one-sided communication, though we have not found it to be faster.
-
A paper critiquing HPCG was discussed: Marjanović, Gracia, and Glass, Performance modeling of the HPCG benchmark.
We encourage interested parties to continue discussion on the hpgmg-forum mailing list.