Relying on computing systems that become increasingly complex is difficult: with many factors potentially affecting the result of a computation or its properties, understanding where problems appear and fixing them is a challenging proposition. Typically, the process of finding solutions is driven by trial and error or by experience-based insights.

In this dissertation, I examine the idea of using provenance metadata (the set of elements that have contributed to the existence of a piece of data, together with their relationships) instead. I show that considering provenance a primitive of computation enables the exploration of system behaviour, targeting both retrospective analysis (root cause analysis, performance tuning) and hypothetical scenarios (what-if questions). In this context, provenance can be used as part of feedback loops, with a double purpose: building software that is able to adapt for meeting certain quality and performance targets (semi-automated tuning) and enabling human operators to exert high-level runtime control with limited previous knowledge of a system's internal architecture.

My contributions towards this goal are threefold: providing low-level mechanisms for meaningful provenance collection considering OS-level resource multiplexing, proving that such provenance data can be used in inferences about application behaviour and generalising this to a set of primitives necessary for fine-grained provenance disclosure in a wider context.

To derive such primitives in a bottom-up manner, I first present Resourceful, a framework that enables capturing OS-level measurements in the context of application activities. It is the contextualisation that allows tying the measurements to provenance in a meaningful way, and I look at a number of use-cases in understanding application performance. This also provides a good setup for evaluating the impact and overheads of fine-grained provenance collection.

I then show that the collected data enables new ways of understanding performance variation by attributing it to specific components within a system. The resulting set of tools, Soroban, gives developers and operation engineers a principled way of examining the impact of various configuration, OS and virtualization parameters on application behaviour.

Finally, I consider how this supports the idea that provenance should be disclosed at application level and discuss why such disclosure is necessary for enabling the use of collected metadata efficiently and at a granularity which is meaningful in relation to application semantics.