Towards Evaluation Mechanisms for Runtime Adaptivity: from Case Studies to Metrics