Using different levels of abstraction for parallel program debugging

Debugging parallel and distributed programs is sometimes a very hard job. For that reason lots of tools have been developed that support the user during error detection and performance analysis. Many of these tools reduce the complexity by providing graphical representations for investigations, some of them even with connections to the original source code. However, most of them operate on a fixed level of abstraction, which can be problematic in several cases. In contrast, the MAD environment is a tool set which consists of several specialized modules for dedicated activities of parallel program debugging on different levels of abstraction. Starting with a very general kind of program representation-the event graph-the user can easily switch between different abstractions. Examples for using the tool set and cases where different levels of abstraction are necessary are shown in this paper.