Causal models for debugging and control in cloud computing

Two challenges in the theory and practice of cloud computing are: (1) smart control (allocation) of resources under exploration constraints, on time-varying systems, and (2) understanding and debugging of the performance of complex systems that involve e.g. virtualization. In this paper, we examine how these challenges can be approached using causal models. For challenge (1) we investigate how to infer and use causal models and selection diagrams to design and integrate experiments in a principled way, as well as to cope with partially varying systems. For challenge (2) we examine how to formalize performance attribution and debugging questions by counterfactual probabilities, and how to approximately answer them based on inferred (non-deterministic) graphical causal models.