SCCS Colloquium - May 29, 2019
|Date:||May 29, 2019|
|Time:||15:00 - 15:30|
Shreyas Shenoy: Towards Non-blocking Combination Schemes in the Sparse Grid Combination Technique
This is a Master's thesis submission talk. Shreyas is advised by Michael Obersteiner.
Despite the utilization of High Performance Computing (HPC) to solve high-dimensional partial differential equation (PDEs), the extent to which these problems are solvable in considerable amount of time is still restricted by the so called curse of dimensionality. One promising method that strives to mitigate this issue is the sparse grid combination technique. A sparse grid is a suitable approximation of the regular grid, capable of being decomposed further into different coarse and anisotropic computational grids of lower resolutions, called component grids. This enables dual levels of parallelism, as the component grids can be computed in parallel, completely independent from each other. Subsequently the computation on each component grid can also be subject to parallelism. However, certain time-dependent problems, in order to fulfill their stability and convergence riteria. necessitate the combination of component grid solution into sparse grid at regular intervals. This in turn, reintroduces the need for global synchronization and communication. Limiting the extent of scalability of this technique.
In this work, we introduce non-blocking/asynchronous combination techniques, where the combination step is also carried out in parallel with other computational steps, thus striving to eliminate the time lost, during global communication and synchronization. The credibility of this technique has been tested out on the Advection-Diffusion problem and GENE, a gyrokinetic simulation of plasma microturbulence in fusion devise. The ensuing speedup was found to be 5 times of those achieved using normal combination technique.
Keywords: Sparse Grid, Combination Technique, GENE