Classical algebraic multigrid using CUDA

by Simon Layton

Mechanical Engineering Department, Boston University


Abstract

We present a preliminary implementation of a classical algebraic multigrid (AMG) using NVIDIA’s CUDA, including verification against Hypre, a well-known and highly regarded open source AMG code. We illustrate some of the challenges and bottlenecks in obtaining high performance, and show different approaches to overcoming these challenges.

Speaker Bio

Simon Layton obtained his Masters in Mechanical Engineering from Boston University in 2011, and a Bachelors in mathematics and computer science from the University of Bristol in 2008. He is currently a PhD candidate under the supervision of Professor Barba at Boston University. During his postgraduate studies, he has worked on GPU-based projects, including the Fast Gauss transform and a CUDA based implementation of the immersed boundary method in fluid dynamics. Currently he is working on a GPU accelerated classical algebraic multigrid, work begun while working as an intern at NVIDIA in Jonathan Cohen’s emerging applications group during the Summer of 2011.