A Distributed Agent-Based Approach for Coupled Problems in Computational Engineering
Jan Hegewald · Digitale Bibliothek Braunschweig (Verbundzentrale Göttingen (VZG)) · 2011
Challenging questions in science and engineering often require to decouple a complex problem and to focus on isolated sub-problems first. The knowledge of those individual solutions can later be combined to obtain the result for the full question. A similar technique is applied in numerical modeling. Here, the software solver for subsets of the coupled problem might already exist and can directly be used. This thesis describes a software environment capable of combining multiple software solvers, the result being a new, combined model. Two important design decisions were crucial at the beginning: First, every sub-model keeps full control of its execution. Second, the source code of the sub-model requires only minimal adaptation. The sub-models choose themselves when to issue communication calls, with no outer synchronisation mechanism required. The coupling of heterogeneous hardware is supported as well as the use of homogeneous compute clusters. Furthermore, the coupling framework allows sub-solvers to be written in different programming languages. Also, each of the sub-models may operate on its own spatial and temporal scales. The next challenge was to allow the potential coupling of thousands software agents, being able to utilise today's petascale hardware. For this purpose, a specific coupling framework was designed and implemented, combining the experiences from the previous work with additions required to cope with the targeted number of coupled sub-models. The large number of interacting models required a much more dynamic approach, where the agents automatically detect their communication partners at runtime. This eliminates the need to explicitly specify the coupling graph a~priori. Agents are allowed to enter (and leave) the simulation at any time, with the coupling graph changing accordingly.