Below you can see the layout of a pixel of MPEC 3.0.
NEC which are located in the Tokyo area. I already had substantial knowledge about compiler optimization techniques when I went to Japan, and I used it to optimize the Linpack Benchmark program.
The internship was very successful, and at the end of the three months I had achieved a new speed world record for Pentium III systems.I worked in the High-Performance Computing Group, and it was the first time that I could work as a full-time scientist over a longer period. I gained a lot through this internship: The Japanese scientists showed me how to do focused and concentrated research on a high level. Interestingly only months before, the Earth Simulator was finished by NEC, which became the world's most powerful supercomputer by far. I could visit it at the end of my stay, and I have put some
private photos of the supercomputer online.
The compilers for the Earth Simulator were written by the research group I had joined over the summer.It was a great honor that I was accepted by such a top research team.
DFG. Considering my limited research experience at that time, that was a huge honor and a significant step forward for me.
The project was pretty big and incorporated several research groups. It was supervised by the CVGPR Group of the University of Mannheim. Basically, the goal was to implement an isotropic nonlinear diffusion filtering on clusters. My task was the optimization of the local memory access pattern and the parallelization with threads to be able to fully exploit the multi-processor nodes of clusters.At the end I had the chance to test the program on
HELICS, the Heidelberg Linux Cluster System, which was brand-new and the world's most powerful Linux cluster at that time.
Chair of Computer Architecture of the University of Mannheim. I designed a basic program that used the time-stamp counter to measure the latency of memory references. The idea was to analyze the execution speed of a sufficiently complex and suitably structured reference pattern to determine the exact structure and speed of the memory hierarchy.
The program I wrote might serve as a basis for a future benchmark suite of the University of Mannheim.