Towards a discipline of performance engineering : lessons learned from stencil kernel benchmarks

High performance computing systems are characterized by a high level of complexity both on their hardware and software side. The hardware has evolved offering a lot of compute power, at the cost of an increasing effort needed to program the systems, whose software stack can be correctly managed only by means of ad-hoc tools. Reproducibility has always been one of the cornerstones of science, but it is highly challenged by the complex ecosystem of software packages that run on HPC platforms, and also by some malpractices in the description of the configurations adopted in the experiments. In this work, we first characterize the factor that affects the reproducibility of experiments in the field of high performance computing and then we define a taxonomy of the experiments and levels of reproducibility that can be achieved, following the guidelines of a framework that is presented. A tool that implements said framework is described and used to conduct Performance Engineering experiments on kernels containing the stencil (structured grids) computational pattern. Due to the trends in architectural complexity of the new compute systems and the complexity of the software that runs on them, the gap between expected and achieved performance is widening. Performance engineering is critical to address such a gap, with its cycle of prediction, reproducible measurement, and optimization. A selection of stencil kernels is first modeled and their performance predicted through a grey box analysis and then compared against the reproducible measurements. The prediction is then used to validate the measured performance and vice-versa, resulting in a "Gold Standard" that draws a path towards a discipline of performance engineering