P16: Scaling Analysis of a Hierarchical Parallelization
of Large Inverse Multiple-Scattering Solutions
SessionPoster Reception
Event Type
ACM Student Research Competition
Poster
Reception
TimeTuesday, November 14th5:15pm -
7pm
LocationFour Seasons Ballroom
DescriptionWe propose a hierarchical parallelization strategy to
improve the scalability of inverse multiple-scattering
solutions. The inverse solver parallelizes the
independent forward solutions corresponding to different
illuminations. For further scaling out on large numbers
of computing nodes, each forward solver parallelizes the
dense and large matrix-vector multiplications
accelerated by the multilevel fast multipole algorithm.
An inverse problem involving a large Shepp-Logan phantom
is solved on up to 1,024 CPU nodes of the Blue Waters
supercomputer in order to demonstrate the strong-scaling
efficiency of the proposed parallelization scheme. The
results show that parallelizing illuminations has almost
perfect scaling efficiency of 95% because of the
independent nature of forward-scattering solutions,
however, parallelization of MLFMA has 73% efficiency due
to MPI communications in MLFMA multiplications.
Nevertheless, the proposed strategy improves granularity
and allows spreading DBIM solutions on large numbers of
nodes.
