Design and Implementation of Parallel Computing Models for Solar Radiation Simulation

Abstract

In order to simulate geographical phenomenon, many complex and high precision models have been developed by scientists. But at most time common hardware and implementation of those computation models are not capable of processing large amounts of data, and the time performance might be unacceptable. Nowadays, the growth in the speed of modern graphics processing units is incredible, and the flops/dollar radio provided by GPU is also growing very fast, which makes large scale GPU clusters gain popularity in the scientific computing community. However, GPU programming and clusters' software deployment and development are associated with a number of challenges.

In this thesis, the geo-science model developed by I. D. Dobreva and M. P. Bishop proposed in A Spatial Temporal, Topographic and Spectral GIS based Solar Radiation Model (SRM) was analyzed. I built a heterogeneous cluster and developed its software framework which could provide powerful computation service for complex geographic models. Time performance and computation accuracy has been analyzed. Issues and challenges such as GPU programming, job balancing and scheduling are addressed.

The SRM application running on this framework can process data fast enough and be able to give researchers rendering images as feedback in a short time, which improved the performance by hundreds of times when compared to the current performance in our available hardware, and the speedup can easily be scaled by adding new machines.