Simplifying the use of event-based systems with context mediation and declarative descriptions

Current trends like the proliferation of sensors or the Internet of Things lead to Cyber-physical Systems (CPSs). In these systems many different components communicate by exchanging events. While events provide a convenient abstraction for handling the high load these systems generate, CPSs are very complex and require expert computer scientists to handle correctly. We realized that one of the primary reasons for this inherent complexity is that events do not carry context. We analyzed the context of events and realized that there are two dimensions: context about the data of an event and context about the event itself. Context about the data includes assumptions like systems of measurement units or the structure of the encoded information that are required to correctly understand the event. Context about the event itself is data that provides additional information to the information carried by the event. For example an event might carry positional data, the additional information could then be the room identifier belonging to this position. Context about the data helps bridge the heterogeneity that CPSs possess. Event producers and consumers may have different assumptions about the data and thus interpret events in different ways. To overcome this gap, we developed the ACTrESS middleware. ACTrESS provides a model to encode interpretation assumptions in an interpretation context. Clients can thus make their assumptions explicit and send them to the middleware, which is then able to mediate between different contexts by transforming events. Through analysis of the provided contexts, ACTrESS can generate transformers, which are dynamically loaded into the system. It does not need to rely on costly operations like reflection. To prove this, we conducted a performance study which shows that in a content-based publish/subscribe system, the overhead introduced by ACTrESS’ transformations is too small to be measurable. Because events do not carry contextual information, expert computer scientists are required to describe situations that are made up of multiple events. The fact that CPSs promise to transform our everyday life (e.g., smart homes) makes this problem even more severe in that most of the target users cannot use CPSs. In this thesis, we developed a declarative language to easily describe situations and a desired reaction. Furthermore, we provide a mechanism to translate this high-level description to executable code. The key idea is that events are contextualized, i.e. our middleware enriches the event with the missing contextual information based on the situation description. The enriched events are then correlated and combined automatically, to ultimately be able to decide if the described situation is fulfilled or not. By generating small computational units, we achieve good parallelization and are able to elegantly scale up and down, which makes our approach particularly suitable for modern cloud architectures. We conducted a usability analysis and performance study. The usability analysis shows that our approach significantly simplifies the definition of reactive behavior in CPS. The performance study shows that the achieved automatic distribution and parallelization incur a small performance cost compared to highly optimized systems like Esper.