A tool for optimizing large scale parallel applications

Developments in the field of supercomputing during the last couple of years showed that new classes of applications (The Grand Challenges) gain in importance that demand for new dimensions in computing power. The development of efficient parallel programs has to be supported in all phases of the program design and coding. This paper introduces a tool that supports the optimization of programs on massively parallel computing systems. The tool is scalable with respect to its implementation and in the way it presents performance data. A major feature contributing to the scalable representation of performance data is the ability to focus measurements on points of interest in the program execution by specifying behavioral attributes. Behavioral attributes are given as thresholds to the results of other measurements. Thus a direct link between results of different measurements can be made which enables the user to link global system behavior to the execution of individual program parts. With regard to the implementation, it is shown how the distributed measurement evaluation and the on-line mode of operation enable the tool to handle the problems of tools for massively parallel computing platforms