Cloud-edge acceleration strategies for discrete event simulation in manufacturing systems

In light of sustainability, this thesis examines GPU and as-a-service strategies to accelerate the dominant category of manufacturing simulation, Discrete Event Simulation, in heterogeneous cloud-edge systems.Initially, cloud-based GPU computing was applied to accelerate three manufacturing factory case simulations of increasing complexity: source-sink, assembly-line balancing and reinforcement learning supply strategy.

The same simulations, at equivalent computational scales and complexities, were subsequently analysed using edge-based GPU computing - consisting of two modes of execution: low-resourced and high-resourced. Speedups spanned between 1.2x and 3.2x, alongside the novel identification and quantification of the non-proportional relationship between computational scale and speedup based on high-level GPU acceleration in the cloud-edge spectrum.

Stemming from this was a time, energy and cost analysis to discern how each metric was intertwined during DES execution. For the first time in the context of industrial manufacturing - the trade-offs between DES speed, energy and cost in the cloud-edge were described, the offset between each cloud-edge domain for each metric was measured, and the optimum cloud-edge domain was identified. The culmination of these contributions was the implementation of DES-as-a-service.

This incorporated workload sharing between cloud and edge, and demonstrated an effective approach to scale up, offload and accelerate simulation - achieving a speedup high of 7.3x.