The HALadapt project is aimed at hiding the complexity of managing a heterogeneous parallel system from an application developer and reducing, through targeted optimizations, for example, the energy consumption or execution time of applications.

Project start: 01.01.2016
Project group: Prof. Dr. Wolfgang Karl
Contact:
Links: Jobs - Theses

Modern heterogeneous parallel systems are an integral part of everyday life in numerous fields of application such as embedded systems or high-performance computing. These different areas of application are linked by a dynamic in the requirements placed on the computing system used. New boundary conditions and optimization goals for the computing system can arise at runtime, for example due to influences from the computing system's environment. This makes a static configuration of the system a complex problem for a user.

The HALadapt runtime system uses self-organization methods to proactively adapt the computing system to new requirements and environmental influences. In particular, the mapping of the applications to be executed as so-called "compute kernels" is proactively adapted to the current and future system states. For this purpose, HALadapt uses a machine learning-based approach that learns rules that try to find an optimizing balance between conflicting optimization goals such as minimizing application runtimes, minimizing energy consumption, minimizing heat generation of the computing units and maximizing system reliability.