MEPHISTO - Metaprogramming for Heterogeneous Distributed Systems
The aim of the project is to make the performance of modern HPC systems optimally exploitable for application programmer, and to achieve high performance on all current high-performance computers without changing the source code. Individual computation nodes are so complex that different memory and parallelism levels must be deliberately taken into account in order to achieve this. The current hardware concepts differ not only from each other, but also require completely different programming models for the application software. In addition, the systems consist of more and more homogeneous or heterogeneous compute devices, whose collabora-tion is crucial for the efficient resolution of major problems that require the overall computing power of the system. Again, there are very different concepts, which are considered to be opti-mal depending on the application and the system. In order to achieve these objectives, data con-tainers and algorithms are made available to enable application programmers to make the best use of the performance potential without special knowledge in parallel programming. These data containers not only automatically distribute the data over the entire high-performance computer, they also enable portable programming of accelerator architectures, such as are now available in heterogeneous data arithmetic nodes.
Project Website
Partners
- Center for Information Service and High Perfprmance Computing at Technische Universität Dresden
- Institut für Informatik, LFE für Kommunikationssysteme und Systemprogrammierung an der Ludwig-Maximilians-Universität (LMU) München
- Helmholtz-Zentrum Dresden-Rossendorf (HZDR)
ZIH-Contact
Dr. Andreas Knüpfer
Project Term
02/2017-01/2020
Funding
BMBF (FKZ 01IH16006A)