Multi-level service monitoring

Nowadays, applications are more and more realized as service compositions, that is, assembling small, independent, accessed on-demand, possibly heterogeneous pieces of code in an easy and flexible way. Cloud computing takes the service abstraction one step further and applies it to new kinds of resources, both hardware and software. On the other hand, service-based applications are way more complex to manage than traditional ones. They, in fact, are naturally structured on multiple, interdependent layers. This means we need a new way to control their performance, which would take into account all the functional and non-functional dependencies and would be able to react to the frequent changes of services. There are many existing runtime service management solutions. However, in the context of cloud computing, they can only be considered partial solutions: some of them intervene at specific levels only, while others do not consider the adaptation process in its totality. This thesis refers to a project in which runtime management is obtained through a multi-level MAPE cycle. The approach consists of four phases that are applied concurrently to multiple layers. The four phases are: monitoring and correlation; analysis of adaptation needs; identification of cross-layer strategies; adaptation enactment. This thesis focuses on the realization of the first phase in the cycle: monitoring. The work proposes an approach in which we extract independent execution data from each layer, and correlate them to obtain a clear and holistic representation of the application’s functional and non-functional behaviour. In addition, the thesis presents two performance visualization tools. The first is real-time, and can show live trends. The second uses historical data to allow the designer to “drill down” to discover the reasons behind past application failures. The entire approach has been tested with various load simulations on a distributed application.