Self-optimization of large scale wildfire simulations

The development of efficient parallel algorithms for large scale wildfire simulations is a challenging research problem because the factors that determine wildfire behavior are complex. These factors make static parallel algorithms inefficient, especially when large number of processors is used because we cannot predict accurately the propagation of the fire and its computational requirements at runtime. In this paper, we propose an Autonomic Runtime Manager (ARM) to dynamically exploit the physics properties of the fire simulation and use them as the basis of our self-optimization algorithm. At each step of the wildfire simulation, the ARM decomposes the computational domain into several natural regions (e.g., burning, unburned, burned) where each region has the same temporal and special characteristics. The number of burning, unburned and burned cells determines the current state of the fire simulation and can then be used to accurately predict the computational power required for each region. By regularly monitoring and analyzing the state of the simulation, and using that to drive the runtime optimization, we can achieve significant performance gains because we can efficiently balance the computational load on each processor. Our experimental results show that the performance of the fire simulation has been improved by 45% when compared with a static portioning algorithm.