The aim of this project is to develop a way to generate efficient code for parallel architectures to simulate carbon nanotube structures. It focuses on shared memory parallel machines, e.g., x86 processors, yet also considers accelerators and distributed memory systems.
Performance models provide significant insight of a target application's scaling abilities. However, the creation usually involves a significant amount of time consuming, manual work. In this project, an automated approach to generate performance models is added to the Scalasca tool. In general, simplicity and ease of use are important goals, as is a good estimate of the scaling behavior of different software components relative to each other.
The tuning of parallel applications is a complex tasks that involves several stake holders. We propose a structured approach to follow in a tuning project that aims at maximizing the benefit for the user of the code. The workflow includes blueprints for steps to take and when to communicate which insight. However, it should always be adopted to the current tuning project at hand. (The description is only available in German.)
In this project, fundamentals of tools and methodologies for the numerical simulation of various domains are explored. It focuses on the widely-used approach of multi-grid solvers. These solvers play an important role in many disciplines and areas of applications. (The description is only available in German.)