SIParCS 2016 - Vinay Ramakrishnaiah

Vinay Ramakrishnaiah, University of Wyoming

Restructuring the multi-resolution approximation for spatial data to reduce the memory footprint and to facilitate scalability

(Slides)  (Recorded Talk)

High-resolution observations of spatial fields over large geographic regions from satellites are available at an increasing rate, and their analysis can lead to new insights. Traditional spatial statistical techniques are computationally infeasible for large data sets due to the need to invert a covariance matrix the size of the data. The recently developed multi-resolution approximation (MRA) algorithm, which expresses a spatial process as a linear combination of basis functions at multiple spatial resolutions within a hierarchical framework, addresses this issue. However, since the number of observations determines the memory footprint during execution, a large amount of system RAM is needed for MRA involving huge spatial data. The solution is to utilize the parallelization provided by MRA at each spatial resolution within the hierarchical framework. By having a distributed memory parallel execution, the memory needed per computing node can be reduced. We propose and analyze two implementations to distributed memory parallel algorithm. The first approach is to divide a given resolution layer into smaller parts and assign those parts to individual computing nodes. The second approach creates sub-trees of the multiple spatial resolution hierarchy, and assigns each sub-tree to a computing node. The alternative approaches are found to reduce the memory footprint of the current implementation and show good scalability for huge spatial data.

Mentors: Dorit Hammerling, Raghu Prasanna Kumar, Rich Loft, CISL