Objective: As multi-core processors become ubiquitous and supercomputers grow ever faster, parallel programs allow scientists to apply massive computational power to their problems. Serial implementations of adaptive spacetime solution methods deliver

1
Parallel Implementation of Adaptive Spacetime
Simulations A. Becker and L. V. Kale,
Department of Computer Science
Objective As multi-core processors become
ubiquitous and supercomputers grow ever faster,
parallel programs allow scientists to apply
massive computational power to their problems.
Serial implementations of adaptive spacetime
solution methods deliver unprecedented resolution
for multi-scale problems parallel
implementations promise exciting breakthroughs in
our ability to simulate materials physics. We are
developing new software infrastructure to realize
this potential. However, the asynchronous
character of the method and highly dynamic mesh
refinement pose significant challenges. Approach
We implement our adaptive spacetime meshing
algorithm in ParFUM, a general framework for
parallel finite element programs. ParFUM handles
the communication and synchronization needed to
divide the mesh across hundreds or thousands of
processors while managing adaptive mesh
operations. Impact Spacetime multiscale
simulations can take days on a single processor.
By running simulations in parallel on
supercomputers, we can reduce execution times to
hours or minutes, allowing scientists to study
problems of much greater physical complexity.
Left Spacetime finite element mesh of crack-tip
wave scattering. Vertical axis represents time
horizontal axes represent space. The propagation
of shock waves is readily apparent from the
pattern of adaptive mesh refinement. Below We
decompose the space mesh into chunks
distributed across many processors. Adaptive mesh
refinement and coarsening operations must operate
across chunk boundaries (shown as heavy lines).
Boundaries must be adjusted and chunks
redistributed to minimize communication and
balance the load between the processors.