Authors: Chidchanok Choksuchat; Chantana Chantrapornchai
Addresses: Department of Computing, Faculty of Science, Silpakorn University, Nakhon Pathom, 73000, Thailand ' Department of Computer Engineering, Faculty of Engineering, Kasetsart University, Bangkok, 10900, Thailand
Abstract: Graphics processing units (GPUs) are the important accelerators in our desktop computer nowadays. There are thousands of processing units that can simultaneously run the scientific program and there are various memory types, with different sizes and access times, which are connected in a hierarchy. However, the GPUs have a much smaller internal memory size than a typical computer, which can be obstacles to handle all of a dataset when performing big data processing. In this paper, we study the use of various memory types: global, texture, constant, and shared memories, in simultaneously searching large Resource Description Framework (RDF) data which are commonly used on the internet to link to the WWW data based on the GPUs. Using suitable memory types and properly managing the data transfer between host and device can lead to a better performance when processing such data. The results show that the parallel search utilises the global memory for storing large texts in 45-Gigabyte RDF data on multiple GPUs and utilises the shared memory for storing multiple keywords and can run about 14 times faster than the sequential search on a low-cost desktop.
Keywords: graphics processing units; GPUs; large RDF; parallel string search.
International Journal of Computational Science and Engineering, 2018 Vol.17 No.4, pp.398 - 410
Available online: 07 Nov 2018 *Full-text access for editors Access for subscribers Purchase this article Comment on this article