Design and proto-typical implementation of an analysis tool interface for a task-based PGAS runtime

To build fast parallel applications, multiple programming models have been developed over the past years. In particular, the Partitioned Global Address Space (PGAS) model has emerged from the traditional shared memory and distributed memory models. The PGAS model offers decoupled synchronization and communication between processes. Recent approaches combine the PGAS model with task parallelism, e.g. DASH, UPC++ and X10. Tools are needed to verify the correctness and analyze the performance of parallel applications. To use such tools, an interface is needed to enable communication between the underlying runtime and the tool itself. In this thesis, the design of a tools interface for the DASH C++ PGAS Framework is explored and evaluated. A modular plugin-based interface infrastructure is developed and implemented to connect external analysis tools. Plugins are used as a wrapper for external analysis tools, which do not offer native support for the developed tools interface. To show the functionality of the interface and the corresponding infrastructure, two analysis tools, namely Temanejo and Extrae are connected to the interface infrastructure. Finally, the interface infrastructure and the plugin to connect the Extrae performance analysis tool are evaluated by using both real-world and microbenchmarks to determine a possible overhead. For the infrastructure, no significant overhead can be measured with both real-world and microbenchmarks. However, with microbenchmarks, a significant overhead can be measured for the Extrae plugin. As the real-world benchmark reveals, the overhead is not relevant in practice, since no significant overhead can be measured here