| Description |
-
Abstract: Multigrid methods have long been used as efficient solvers for a
large class of discrete differential equations. Recently, parallel multigrid solvers have
been primarily implemented within an MPI model, with scalability to thousands of cores.
With manycore GPU architectures gaining popularity, there is an increased desire to extend
multigrid solvers to these architectures. In these ... read moreenvironments, algebraic multigrid
methods can be cumbersome due to their reliance on indirect addressing. In contrast,
structured-grid methods, such as BoxMG, achieve efficiency for heterogeneous problems with
only direct addressing. We will discuss parallelization of BoxMG on a variety of
architectures, and outline the implementation of accelerated BoxMG using OpenCL. By
minimizing the data transfer between the CPU and GPU, we will show that significant
performance improvements can be achieved over serial BoxMG, close to the bounds predicted
by Amdahl's Law.
Thesis (M.S.)--Tufts University, 2012.
Submitted to the Dept. of Mathematics.
Advisor: Scott MacLachlan.
Committee: James Adler, and Christoph Borgers.
Keyword: Mathematics.read less
|
|---|
This object is in collection
