Author

Sandip Sahani

Date of Award

2015

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

Committee Chair

Chao Peng

Committee Member

Ramazan Aygun

Committee Member

Sun-il Kim

Subject(s)

Image processing--Digital techniques, Computer graphics, Real-time programming, Computer algorithms, Weather forecasting

Abstract

In this thesis, we propose a GPU based approach to process large weather datasets. This work contributes to big data research with a parallel approach towards GPU architecture. A GPU-based algorithm is presented to perform 3D Connected Component Labeling algorithm in parallel. We successfully remove data-dependencies embedded between data frames to achieve pixel-level parallelism. Due the extremely large size of datasets, data frames cannot fit into GPU memory or CPU main memory and sometimes even on the hard drive. Frames have to be streamed through the memory hierarchy, partitioned and processed as batches, where each batch is large enought to fit into GPU memory. To merge all batches and reconnect the separated components caused by batch partitioning, we present a batch merging algorithm to extract the information of batch connectivity. The information is stored in a hash table structure on the GPU, which is memory-efficient and supports fast indexing for GPU threads. The merging algorithm do not need to consider all frames of each batch; instead, it needs only the border frames of each batch to extract the information, which makes it possible to merge multiple batches on the GPU. The experiment shows promising results with the implementation in a single GPU. It provides a low cost solution for labeling large weather data. Also, our parallel approach can be seamlessly integrated with general out-of-core techniques. While data streams through hard drive to CPU main memory, this approach could be used as an independent processing component that further streams data into GPU and sends it back to the main data flow for the high-performance computing purpose.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.