Performance prediction for hardware-software configurations is a relevant and practically important problem. With an increasing availability of data in the form of performance measurements, this problem becomes amenable to machine learning, i.e., the data-driven construction of predictive models. In this paper, we propose a learning method that is specifically tailored to the task of performance prediction and takes two important characteristics of this problem into account: (i) prior knowledge in the form of monotonicity constraints, suggesting that certain properties of hard-or software can influence performance only positively or negatively, and (ii) strong differences in the precision and reliability of performance measurements available as training data. We evaluate our method on a real-world dataset from the domain of performance prediction in video games, which we specifically collected for this purpose.