White Rose University Consortium logo
University of Leeds logo University of Sheffield logo York University logo

Working With Incremental Spatial Data During Parallel (GPU) Computation

Chisholm, Robert (2019) Working With Incremental Spatial Data During Parallel (GPU) Computation. PhD thesis, University of Sheffield.

[img]
Preview
Text (Thesis)
Thesis [Corrected Version,2020-01-24] (1).pdf
Available under License Creative Commons Attribution-Noncommercial-No Derivative Works 2.0 UK: England & Wales.

Download (3686Kb) | Preview

Abstract

Central to many complex systems, spatial actors require an awareness of their local environment to enable behaviours such as communication and navigation. Complex system simulations represent this behaviour with Fixed Radius Near Neighbours (FRNN) search. This algorithm allows actors to store data at spatial locations and then query the data structure to find all data stored within a fixed radius of the search origin. The work within this thesis answers the question: What techniques can be used for improving the performance of FRNN searches during complex system simulations on Graphics Processing Units (GPUs)? It is generally agreed that Uniform Spatial Partitioning (USP) is the most suitable data structure for providing FRNN search on GPUs. However, due to the architectural complexities of GPUs, the performance is constrained such that FRNN search remains one of the most expensive common stages between complex systems models. Existing innovations to USP highlight a need to take advantage of recent GPU advances, reducing the levels of divergence and limiting redundant memory accesses as viable routes to improve the performance of FRNN search. This thesis addresses these with three separate optimisations that can be used simultaneously. Experiments have assessed the impact of optimisations to the general case of FRNN search found within complex system simulations and demonstrated their impact in practice when applied to full complex system models. Results presented show the performance of the construction and query stages of FRNN search can be improved by over 2x and 1.3x respectively. These improvements allow complex system simulations to be executed faster, enabling increases in scale and model complexity.

Item Type: Thesis (PhD)
Keywords: GPU,CUDA,spatial partitioning,complex systems,agent based
Academic Units: The University of Sheffield > Faculty of Engineering (Sheffield) > Computer Science (Sheffield)
The University of Sheffield > Faculty of Science (Sheffield) > Computer Science (Sheffield)
Identification Number/EthosID: uk.bl.ethos.798105
Depositing User: Robert Chisholm
Date Deposited: 04 Feb 2020 11:13
Last Modified: 01 Mar 2020 10:53
URI: http://etheses.whiterose.ac.uk/id/eprint/25853

You do not need to contact us to get a copy of this thesis. Please use the 'Download' link(s) above to get a copy.
You can contact us about this thesis. If you need to make a general enquiry, please see the Contact us page.

Actions (repository staff only: login required)