GPGPU parallelization of self-calibrating agent-based influenza outbreak simulation

Abstract

Agent-based simulations of influenza spread are useful for decision making during public health emergencies. During such emergencies, decisions are required in cycles of less than day, and agent-based models should be adapted to support such decisions. The most important considerations for model adaptation are fast calibration of the model, low computational complexity as the population size is scaled up, and dependability of the results with low replication quantity. In previous work, we presented a self-calibrating model for agent-based influenza simulations. We now investigate whether general-purpose GPU computation is effective at accelerating the processing of this model to support health policy decision-making for pandemic and seasonal strains of the virus. The results of this paper indicate that a speedup of 94.3x is obtained with GPU algorithms for simulation sizes of 50 million people. Our GPU implementation scales linearly in the number of people which makes it a good choice for real-time decision support.

Towards a Framework for High-Performance Simulation of Livestock Disease Outbreak: A Case Study of Spread of African Swine Fever in Vietnam

“…Testing approaches on a cluster that has GPU computing nodes indicates that the execution on GPUs may be sped up 7.4 to 11.7 times compared to the CPU run. In order to speed up an influenza propagation agent-based simulation, Holvenstot et al [ 7 ] utilized standard GPU devices. Experimental findings demonstrate that a GPU implementation is far quicker than a multi-threaded CPU implementation.…”

“…The study of applying computer technologies to decision making on disease control in livestock has been conducted by many research groups around the world, mainly focusing on building toolkits to identify characteristic epidemiological features [ 2 , 3 , 4 ] (e.g., use of radio-frequency identification (RFID) tags, surveillance cameras, or infrared thermometers) or programs that simulate the direction and extent of spread [ 5 , 6 , 7 , 8 ] of each disease on each type of livestock. In the second vein, the study of infectious disease transmission from a computational science perspective often occurs at the following three levels: (1) modeling of virus reproduction and deformation (modeling shape-shifting viruses) [ 9 , 10 ]; (2) modeling the immune system focuses on human or animal subjects [ 11 , 12 , 13 ]; (3) modeling disease spread developed for nearly a century, investigating macro factors such as inter-ethnic spread division of the population (city, district, and region) within a country and between countries [ 14 , 15 , 16 ].…”

A service‐based framework for building and executing epidemic simulation applications in the cloud

“…The techniques are tested on a cluster whose compute nodes are equipped with GPUs, and experimental results show that the execution on GPUs can achieve 7.4 to 11.7 times speedup over the execution on CPUs. Holvenstot et al used general‐purpose GPU devices to accelerate an agent‐based simulation of influenza spread. Experimental results show that the GPU implementation achieves significantly greater speedups than a multi‐threaded CPU implementation.…”

“…To summarize, some of the related work for executing epidemic simulations does not exploit clouds and their associated parallelization, cost, and agility benefits . Other related work exploits clouds but lacks support for elasticity, multi‐cloud deployment, fault tolerance, or required facilities beyond bag‐of‐tasks execution .…”