Andrey Vladimirov
(Colfax Research)
Intel Many Integrated Core (MIC) architecture is a computing platform
for energy-efficient and accelerated execution of highly parallel
applications. Unlike General Purpose Graphics Processing Units
(GPGPUs), MIC-based Intel Xeon Phi coprocessors can execute
applications compiled from the same source code in C/C++/Fortran as
the CPU platform, which may be of crucial importance for some legacy
scientific applications. Furthermore, optimization for the MIC
architecture usually requires the same strategies as optimization for
general-purpose multicore CPUs. As a consequence, the effort of
porting a code to Xeon Phi may yield a better performing application
for the CPU. This trend of "double rewards" was apparent in a case
study that I will demonstrate in this talk, where a calculation of
cosmic dust heating in the interstellar medium was adapted for the MIC
architecture. Through code optimization and the use of Intel Xeon Phi
coprocessors, we were able to shorten the duration of a data-driven
Bayesian analysis of Galactic structure parameters from years to
weeks. I will outline the methods used in the optimization process,
which are representative of a general class of physics problems with
voxelized simulation spaces, dense linear algebra, discretized and
analytically fitted functional dependencies of physical quantities.