(PhD01) Modelling of Dynamic Network Objects: New Approaches and Adaptation Challenges for Future HPC Systems
TimeMonday, June 25th1:05pm - 1:09pm
DescriptionModelling complex dynamic systems is a research topic for many years and applied to design, automatization, monitoring, control, modelling and simulation in many domains such as mechanical engineering, coal mining, metallurgy. The high complexity of the simulation model is caused by their nonlinearity, spatial distribution and concurrence of the process parameters. Furthermore, they typically grow large and are highly time variant . As a result of their complexity there are still sequential solvers used in practice that cannot utilise at all modern or future HPC system architectures and limit significantly the problem size and dynamism that can be addressed.
As the existing solvers have been designed with homogeneous hardware and sequential processing in mind they cannot be easily parallelized. This applies in particular for heterogeneous multi-processing systems.
Within this PhD the goal is to apply new numerical methods to this problem space and target also for an increased abstraction from the underlying hardware by utilising frameworks such as AnyDSL .
The chosen use case are different models of coal mine air ventilation systems. Such systems can be represented as dynamic network objects (DNO), allowing for a representation as graphs with different complexity levels. More specifically these are DNOs with concentrated or distributed parameters.
It is planned to base the parallel solver on block difference numerical methods (BDM). These methods allow to compute more than one simulation point per each iteration step (e.g. 4 points, BDM4). The first experiments with sequential BDM-solvers show high accuracy and convergence of these methods, less amount of computational steps, and good “accuracy-speed” relation, what brings a good reason for further investigation in the direction of parallelization. An interesting research point is also a parallelization on the graph level, i.e. granularity of the parallel processes and their mapping to the physical compute resources (different parallelization levels).
In order to achieve the anticipated simplicity it is proposed to express this algorithm as a domain specific language (DSL). The goal is to separate description (in clear form for the end-user, an expert in the domain) and the adaptation to the specific architecture (which should be done by some hardware expert). As a starting point the AnyDSL framework will be investigated.
The poster describes the simulation problem for complex dynamic systems, HPC-specific model adaptation challenges, and the basic concept for the proposed approach using BDM and DSL. There are also first experimental results (for sequential BDM-solvers) and further planned research steps.