Keywords: interactive HPC, sliding window, computational fluid dynamics.

Recent advances in supercomputing, especially the latest exascale initiative, have opened the door to more and more application domains, with yet more and more problems coming into the range of being feasible on massive parallel systems. While most of those codes are (still) running in batch mode, an interactive processing of the underlying simulations of physical phenomena would be advantageous. Coupling a simulation back end, running on a supercomputer or compute cluster, to a visual front end for interaction, often referred to as computational steering [1], allows users to experience immediate feedback concerning the effects of changes in the geometric model, boundary conditions, or even different algorithms used for the problem solution.

While such an interactive processing is doubtlessly beneficial for many engineering applications and optimisation or design problems, it nevertheless entails many new challenges that have to be solved by application programmers. One of those challenges relates to the fast and efficient data transfer between front- and back- end in the case of highly resolved computational domains. Latency is physics, bandwidth is money but even with sophisticated network topologies and interconnects a real time transfer of several gigabytes per second is unrealistic and hinders interactive computing. Instead of transmitting all data for visualisation (typically also involving filtering) on the front end, users should be able to decide at what resolution and which parts of the domain have to be sent to the back end.

Therefore, we present an approach based on a sliding window technique where users can select a certain region, the window, of the entire computational domain to be transmitted to or visualised on the front end. The key feature of this approach is to keep a constant bandwidth all the time, i.e. to transmit the same amount of data independent of the size and position of the window. Hence, for `large' windows filtering of the data points inside that region must be done while for `small' windows it might match the available resolution given by the discretisation of the computational domain. This allows users to seamlessly zoom in or out of the computational domain in order to concentrate either on specific details such as the thermal flow inside a building or on a more global picture such as the flow around a whole city for pollutant transportation. To support this approach, a distributed hierarchical data structure [2,3] has been implemented that allows for both a simple parallel processing and grid migration as required within dynamic load balancing.

In the paper, we discuss the sliding window data transfer in HPC simulations and its benefits for interactive computing scenarios. Furthermore, we demonstrate the usability of this approach for steering a multi-scale fluid flow simulation running on a supercomputer.

1

J.D. Mulder, J.J. van Wijk and R. van Liere, "A survey of computational steering environments", Future Generation Computer Systems, vol. 13, 1998.

2

J. Frisch, R.-P. Mundani and E. Rank, "Adaptive data structure management for grid based simulations in engineering applications", In Proc. of the 8th Int. Conf. on Scientific Computing, 2011.

3

V. Varduhn, R.-P. Mundani and E. Rank, "A framework for parallel numerical simulations on multi-scale geometries", In Proc. of the 11th Int. Symposium on Parallel and Distributed Computing, 2012.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £40 +P&P)