TIME: 3:30 PM

LOCATION: GMCS 314

SPEAKER: Hans Johansen, Lawrence Berkeley National Laboratory (Applied Mathematics and Computational Research Division)

ABSTRACT: Broadly speaking, computational science is about getting the best answer for a given amount of computing resources. But today’s computing resources include CPUs, GPUs, and supercomputers with 1000’s of nodes, and algorithms perform very differently across those scales. In this talk I will make the case for co-design of methods that combine algorithms and software running on the latest hardware to target a given science problem. For example, embedded boundary (also known as “cut cell”) methods for solving partial differential equations can significantly reduce costs for grid generation and computation in complex and moving domains; but EB methods are harder numerically and highly irregular on GPUs. Another example is time integrators, which must have accuracy and stability; they can also be serial bottlenecks and have poor scaling. For this, recent advances combining parallel-in-time and splitting methods yield both higher accuracy and parallelism. Lastly, to take advantage of GPU hardware we have developed a C++ library that implements higher-order methods, as well as multi-material domains, mapped multi-block grids, and adaptive mesh refinement. I’ll provide some demonstrations of how these fit together for different model problems, and make the argument that higher-order methods produce better answers while being more efficient at all scales. This is joint work with Nate Overton-Katz (AMD), Rochi Chowdhury (LBNL), and Stephen Guzik (UVa).

BIO: Dr. Hans Johansen is a computational researcher at Lawrence Berkeley National Laboratory, specializing in numerical discretizations and algorithms for large-scale scientific computing. He focuses on computational science for biology, electronics, and climate, as well as HPC software technologies, and application performance and portability. He has over 30 years of experience in applied math, computational research, IT consulting, startups, and financial services technology.

HOST: Christopher Paolini