A High Performance, Agent-Based Simulation of Old World Screwworm Fly Lifecycle and Dispersal using a Graphics Processing Unit (GPU) Platform

Abstract

The Old World Screwworm Fly (OWSWF), 'Chrysomya bezziana', is an insect pest that is endemic to the tropical regions of Asia, the Middle East and Africa. The insect reproduces by laying its eggs in open wounds and mucus membranes of warm blooded mammals. Upon the hatching, the OWSWF larvae eat the living flesh of the host animal, causing injury, secondary infections and in extreme cases death. If this pest was introduced to the Australian mainland, it could have a devastating impact on the livestock industries within the northern regions of Australia. This work builds upon the existing research surrounding the OWSWF biological lifecycle and dispersal characteristics by developing a national-scale, high-resolution, agent-based model capable of simulating an invasion of Australia by the OWSWF. The challenge in designing large scale high-resolution models to run on personal computers is addressing performance issues. We face this challenge by making use of 'Graphics Processing Unit (GPU)' technologies, based around NVidia's 'Compute Unified Device Architecture (CUDA)', to simulate the lifecycle and dispersal of the OWSWF at the individual insect and cohort levels. This model combines agent-based logic, for simulating the OWSWF's lifecycle, with an efficient cellular-automata system to capture the spatial aspects of the OWSWF population's dispersal during a simulated invasion. The lifecycle and dispersal simulation is supported by an efficient system of main memory management which integrates bio-climatic data from a standard database management system for use within the model. The scheme adopted breaks this agent logic down into GPU-based functions, known as kernels, and uses the well-developed heterogeneous programming approach to distribute processing tasks between the 'Central Processing Unit' (CPU) of the host machine and the CUDA device. Analysis of the performance of the CUDA implementation reveals significant improvement in execution time when compared to an equivalent CPU-only based implementation, with results showing that the CUDA implementation's processing efficiency scales up well as the number of agents within the simulation increase.