Cost-Effective Hardware-Aware Machine Learning to Mitigate Memory and Workload Bottlenecks in Parallel Architectures.