With ubiquitous parallel architectures, the importance of optimally distributed and thereby balanced work is unprecedented. To tackle this challenge, graph partitioning algorithms have been successfully applied in various application areas. However, there is a mismatch between solutions found by classic graph partitioning and the behavior of many real hardware systems. Graph partitioning assumes that individual vertex weights add up to partition weights (here, referred to as linear graph partitioning). This implies that performance scales linearly with the number of tasks. In reality, performance does usually not scale linearly with the amount of work due to contention on various resources. We address this mismatch with our novel penalized graph partitioning approach in this paper. Furthermore, we experimentally evaluate the applicability and scalability of our method.