Offloading computation to the GPU is crucial to leverage many of today’s supercomputers. We expect the GPU port of an application to outperform the pure CPU implementation, but is this always true? Simple benchmarking only allows us to take a limited number of samples from a vast space of execution configurations and can, therefore, deliver only a fragmented answer. To answer the question systematically, even for individual application kernels, we propose a semi-automatic toolchain based on differential performance modeling and intuitive visualizations. We combine empirical performance models based on unified CPU–GPU profiles with hardware characteristics to derive differential performance models that can be easily compared across device types. In four case studies, we demonstrate how our toolchain pinpoints scaling issues in GPU ports, guides performance improvements, and identifies execution configurations with superior performance.

Empirical performance modeling is a proven instrument to analyze the scaling behavior of HPC applications. Using a set of smaller-scale experiments, it can provide important insights into application behavior at larger scales. Extra-P is an empirical modeling tool that applies linear regression to automatically generate human-readable performance models. Similar to other regression-based modeling techniques, the accuracy of the models created by Extra-P decreases as the amount of noise in the underlying data increases. This is why the performance variability observed in many contemporary systems can become a serious challenge. In this paper, we introduce a novel adaptive modeling approach that makes Extra-P more noise resilient, exploiting the ability of deep neural networks to discover the effects of numerical parameters, such as the number of processes or the problem size, on performance when dealing with noisy measurements. Using synthetic analysis and data from three different case studies, we demonstrate that our solution improves the model accuracy at high noise levels by up to 25% while increasing their predictive power by about 15%.