Parallel execution and scriptability in micromagnetic simulations

Download

PDF - Pre print Download (115Kb)

Description/Abstract

We demonstrate the feasibility of an ‘encapsulated parallelism’ approach towards micromagnetic
simulations that combines offering a high degree of flexibility to the user with the efficient utilization
of parallel computing resources.
While parallelization is obviously desirable to address the high numerical effort required for re-
alistic micromagnetic simulations through utilizing now widely available multiprocessor systems
(including desktop multicore CPUs and computing clusters), conventional approaches towards par-
allelization impose strong restrictions on the structure of programs: numerical operations have to be
executed across all processors in a synchronized fashion. This means that, from the user’s perspec-
tive, either the structure of the entire simulation is rigidly defined from the beginning and cannot be
adjusted easily, or making modifications to the computation sequence requires advanced knowledge
in parallel programming.
We explain how this dilemma is resolved in the Nmag simulation package in such a way that the user
can utilize without any additional effort on his side both the computational power of multiple CPUs
and the flexibility to tailor execution sequences for specific problems: simulation scripts written for
single processor machines can just as well be executed on parallel machines and behave in precisely
the same way, up to increased speed. We provide a simple instructive magnetic resonance simulation
example that demonstrates utilizing both custom execution sequences and parallelism at the same
time. Furthermore, we show that this strategy of encapsulating parallelism even allows to benefit
from speed gains through parallel execution in simulations controlled by interactive commands given
at a command line interface.