Fault Tolerant Computation of Hyperbolic Partial Differential Equations with the Sparse Grid Combination Technique [article]

Thomas Brendan Harding, University, The Australian National, University, The Australian National
As the computing power of supercomputers continues to increase exponentially the mean time between failures (MTBF) is decreasing. Checkpoint-restart has historically been the method of choice for recovering from failures. However, such methods become increasingly inefficient as the time required to complete a checkpoint-restart cycle approaches the MTBF. There is therefore a need to explore different ways of making computations fault tolerant. This thesis studies generalisations of the sparse
more » ... id combination technique with the goal of developing and analysing a holistic approach to the fault tolerant computation of partial differential equations (PDEs). Sparse grids allow one to reduce the computational complexity of high dimensional problems with only small loss of accuracy. A drawback is the need to perform computations with a hierarchical basis rather than a traditional nodal basis. We survey classical error estimates for sparse grid interpolation and extend results to functions which are non-zero on the boundary. The combination technique approximates sparse grid solutions via a sum of many coarse approximations which need not be computed with a hierarchical basis. Study of the combination technique often assumes that approximations satisfy an error splitting formula. We adapt classical error splitting results to our slightly different convention of combination level. Literature on the application of the combination technique to hyperbolic PDEs is scarce, particularly when solved with explicit finite difference methods. We show a particular family of finite difference discretisations for the advection equation solved via the method of lines has solutions which satisfy an error splitting formula. As a consequence, classical error splitting based estimates are readily applied to finite difference solutions of many hyperbolic PDEs. Our analysis also reveals how repeated combinations throughout the computation leads to a reduction in approximation error. Generalisations of the combination technique are studied a [...]
doi:10.25911/5d6513f948824 fatcat:sisrsdlwwvc5hp66sg6q2js62i