Accelerating the 'Fields' Package: Using Parallel Linear Algebra Software

07/31/2014 - 11:15am
Mesa Lab Main Seminar Room

Isaac Lyngaas portrait

Isaac Lyngaas, SIParCS Intern
(South Dakota State University)

Statistical packages such as the ‘Fields’ package in R can benefit from using  multicore systems and accelerated hardware such as GPUs and Xeon Phis. Performance of linear algebra operations such as those used in spatial statistics computations in ‘Fields’  can be accelerated by new accelerator hardware. A survey  is completed for  several multicore and accelerator based linear algebra packages, including  the commercial package CULA,  and PLASMA  and MAGMA both from the University of Tennessee, to determine the possible speedup.  For double precision matrices of size N=5000 or greater, we find that computing a Cholesky decomposition using CULA we can get at least an 18 times speedup over conventional R.  This method uses a single Nvidia Tesla M270Q GPU within Caldera. 

Video Presentation