P09: Adaptive Multistep Predictor for Accelerating
Dynamic Implicit Finite-Element Simulations
SessionPoster Reception
Event Type
ACM Student Research Competition
Poster
Reception
TimeTuesday, November 14th5:15pm -
7pm
LocationFour Seasons Ballroom
DescriptionWe develop an adaptive multistep predictor for
accelerating memory bandwidth-bound dynamic implicit
finite-element simulations. We predict the solutions for
future time steps adaptively using highly-efficient
matrix-vector product kernels with multiple right-hand
sides to reduce the number of iterations required in the
solver. By applying the method to a conjugate gradient
solver with 3 x 3 block Jacobi preconditioning, we were
able to achieve a 42% speedup on a Skylake-SP Xeon Gold
cluster for a typical earthquake ground motion problem.
As the method enables the number of iterations, and thus
the communication frequency, to be reduced, the
developed solver was able to attain high size-up
scalability: 80.6% up to 32,768 compute nodes on the K
computer. The developed predictor can also be applied to
other iterative solvers and is thus expected to be
useful for wide range of dynamic implicit finite-element
simulations.
