High Performance Computing with Remotely Sensed Spatial Big Data Using the Many-core Graphics Processing Unit (GPU)
Author | : Feng Ni (Ph.D.) |
Publisher | : |
Total Pages | : 236 |
Release | : 2016 |
ISBN-10 | : OCLC:1038717021 |
ISBN-13 | : |
Rating | : 4/5 ( Downloads) |
Download or read book High Performance Computing with Remotely Sensed Spatial Big Data Using the Many-core Graphics Processing Unit (GPU) written by Feng Ni (Ph.D.) and published by . This book was released on 2016 with total page 236 pages. Available in PDF, EPUB and Kindle. Book excerpt: The advances in remote sensing technologies have led to the dramatic enlargement of the spatial data repositories and the formation of remotely sensed Spatial Big Data, which include hyperspectral imagery with hundreds of spectral bands, hyper-spatial imagery with sub-meter pixel sizes, and LiDAR (Light Detection And Ranging) data with dense 3D point cloud, and massive waveform curves. The traditional serial computing paradigm became extremely or prohibitively time-consuming when involving complex spatial analysis algorithms and large volumes of remotely sensed Spatial Big Data. High-performance computing (HPC) based on parallel processors, especially the many-core Graphics Processing Unit (GPU), is beginning to be utilized in many research areas involving data-intensive and computationally complex tasks. However, the GPU-based parallel processing of remotely sensed Spatial Big Data is still at its infantry. Most of the current research is mostly based on ad hoc parallelization solutions, and the comprehensive parallelization strategy for different kinds of spatial analysis is still missing. The size of the data that can be processed in these studies was still relatively small, not considered to be real Spatial Big Data. To overcome these limitations, a comprehensive parallelization strategy for different kinds of spatial analysis with remotely sensed spatial data of different sizes and types was proposed in this research. Specifically, data partitioning strategies inside GPU for the three major types of spatial analysis including local, focal and zonal analysis were designed to utilize the parallel computing power of the many-core GPU; data partitioning strategies outside GPU were designed to break the resource limitation of the many-core GPU; data referencing approaches were provided to efficiently partition regularly spaced and irregularly spaced spatial data. Case studies involving different types of spatial analysis and different types of remotely sensed data were then conducted to demonstrate the algorithm implementation with the proposed parallelization strategy. By running the developed parallel algorithms on a many-core GPU, the computational efficiency of each case study was greatly improved with a computing speedup ranging from 10X to 40X over the serial algorithms. The parallel programs are able to handle data with a size much bigger than the size of the GPU memory, as long as it can be stored on the local hard disks. While bigger data were used, the computing speedup of parallel algorithms was increased up to 72X, which suggest that large datasets like Spatial Big Data should be used for parallel processing to take full advantage of the computing power of the many-core GPU.