Performance modeling for runtime kernel adaptation: A case study on infectious disease simulation

In many large-scale scientific applications, there may be a compute intensive kernel that largely determines the overall performance of the application. Sometimes algorithmic variations of the kernel may be available and a performance benefit can then be gained by choosing the optimal kernel at runtime. However, it is sometimes difficult to choose the most efficient kernel as the kernel algorithms have varying performance under different execution conditions. This paper shows how to construct a set of performance models to explore and analyze the bottleneck of an application. Furthermore, based on the performance models, a theoretical method is proposed to guide the kernel adaptation at runtime. A component-based large-scale infectious disease simulation is used to illustrate the method. The performance models of the different kernels are validated by a range of experiments. The use of runtime kernel adaptation shows a significant performance gain.