Daniel Ritter

During his BGCE studies Daniel Ritter focused on courses in simulation, optimization and control theory. Besides his master studies he was working student at infoteam Software GmbH, a company that designs software for programmable logical controllers (PLCs) and later teaching assistant who leads a class of 20 undergraduate students in "Algorithms and Data Structures 3". This position was part of his BGCE soft skills program ("Supervised Teaching"). After the second semester he took a half year internship at Siemens Information Systems in Bangalore, India. Afterwards, he wrote his Master Thesis at the Chair for System Simulation in Erlangen. The title was "A Fast Multigrid Solver for Molecular Dynamics on the Cell Broadband Engine". After he graduated from BGCE in spring 2008, he worked for AREVA NP, as a programmer in the field of thermohydraulic simulation of a nuclear reactor core, before he started his Ph.D. studies at the chair for system simulation in October 2008 (see his homepage). Current research is going on in the field of hierarchical coarsened grids.

Master's Thesis:
A Fast Multigrid Solver for Molecular Dynamics on the Cell Broadband Engine

Figure 1: Overview CBE

The STI Cell Broadband Engine (CBE) is a revolutionary new multicore architecture. It is quite exotic compared to other architectures on the market due to its asymmetric layout. Besides powering video games in Sony's Playstation 3 it is especially interesting for scientific computing purposes because of its high floating point performance and memory bandwidth. The source of this high performance is are the eight so-called SPE, which are vector processing units with 256 kB of local store (LS) each connected by a high-speed ring bus (EIB). However, code optimization techniques have to be applied to utilize those specialized cores. A schematic view of the CBE is given in figure 1.

Figure 2: 3-level hierarchical example grid

Within the thesis, a common problem from the field of molecular dynamics was the subject of interest: Poisson's equation on an unbound domain. The infinite size of the domain is the main challenge in the computational treatment: To deal with it in a finite machine, a hierarchical coarsened grid approach was taken into account. A coarsened three-level 2D grid is shown in figure 2. For solving Poisson's equation, a finite volume discretization was done on the interfaces, while within one grid level a standard 7-point stencil was used. This hierarchical grid approach enforces the usage of a fast multi grid solver. This solver was implemented as parallel program on the CBE. A number of code optimization techniques were applied in the development of the algorithm. The output was an efficient solver that makes use of the capabilities of the special architecture. Test runs confirmed that the limiting factor in execution speed is the maximum memory bandwidth of the CBE.

Material