IMAGe Brown Bag Seminar

07/09/2015 - 12:00pm
ML- Directors Conference Room
John Paige

Accelerating the 'fields' spatial statistics package in R: Incorporating MAGMA's Cholesky

John Paige, Lawrence Berkeley National  Laboratory


The `Kriging’ algorithm, central to spatial statistics, is O(n^3) in computation time for n observations.  The biggest computational burden lies in the required Cholesky decomposition, which is also O(n^3), and is associated with the evaluation of the data likelihood function given the spatial model covariance parameters.  In this talk, I introduce the `fields’ package, a freely available R package used for spatial statistics, and discuss the results of accelerating its Cholesky decomposition using the Matrix Algebra on GPU and Multicore Architectures (MAGMA) library.  Timing results are given for both MAGMA’s single and double precision Cholesky decompositions using Caldera computational nodes on the Yellowstone supercomputing environment and a mid-2014 MacBook Pro laptop with stock GPU.  In addition, I show that reducing the precision of the Cholesky decomposition does not, in general, significantly affect the likelihood maximization accuracy, yet can substantially improve performance on the mid-2014 MacBook Pro.

Thursday, July 9, 2015

Time:   12:00PM

Location: Mesa Lab Directors Conference Room