Computational & Technology Resources
an online resource for computational,
engineering & technology publications
Civil-Comp Proceedings
ISSN 1759-3433
CCP: 94
PROCEEDINGS OF THE SEVENTH INTERNATIONAL CONFERENCE ON ENGINEERING COMPUTATIONAL TECHNOLOGY
Edited by:
Paper 14

A Multi-Scale Coupling Scheme on Heterogeneous Supercomputers

S. Roller and H. Klimach

German Research School for Simulation Sciences GmbH and RWTH Aachen, Germany

Full Bibliographic Reference for this paper
S. Roller, H. Klimach, "A Multi-Scale Coupling Scheme on Heterogeneous Supercomputers", in , (Editors), "Proceedings of the Seventh International Conference on Engineering Computational Technology", Civil-Comp Press, Stirlingshire, UK, Paper 14, 2010. doi:10.4203/ccp.94.14
Keywords: coupling, multi-scale aeroacoustics, multi-physics, multi-architecture, parallelisation, supercomputing.

Summary
This paper outlines a parallel coupling mechanism for domains of continuum media. The coupling is based on discrete points and was developed in Reference [1]. Restricting the coupling to discrete points simplifies many geometrical operations and allows complex domain configurations without further ado. Independence in the temporal dimension of neighbouring domains is obtained using a Taylor series expansion from the current state of the explicit time marching schemes considered in this work. The state at the discrete points, required for the coupling, are obtained by means of the method used in each domain. Thus the order of the numerical approximation in space and time can be preserved across domain interfaces. A coupling mechanism for domains of different discretization, models and numerical methods, as presented allows the direct simulation of multi scale problems such as aero-acoustics. In such applications a high computational effort is required to solve each domain in three dimensions. A parallel algorithm is therefore essential for the simulation of any real world examples. Clock cycle frequency in modern chips has reached a practicable limit at around 3 GHz prohibiting a simple sequential execution speed increasement in the forseeable future. This led to an increased variance in the available super computing architectures and new strategies to still enhance the execution speeds. With the help of PACX-MPI [2] it is possible to spread the computation of the heterogeneous simulation across multiple machines with different features. This is used to map the various numerical requirements in the subdomains of the problem to appropiate computational architectures. The used systems are the NEC-SX9 and a Intel Nehalem cluster, which provide quite contrary features. Thus the different kernel in the various domains perform very differently on the two systems and ranges from 40% to less than 1% of the peak performances. With the possibility to pick the architecture for each domain, the most efficient execution platform can be used for each, allowing a cheaper computation of the problems with multiple scales. This is even possible with a comparable weak interconnect between the used machines, as the communication between domains on different scales is dictated by the coarser scale and thus a reduced communication effort in comparison to the necessary communication within the domain on the finer scale.

References
1
J. Utzmann, T. Schwartzkopff, M. Dumbser, C. Munz, "Heterogeneous domain decomposition for computational aeroacoustics", AIAA journal, 44(10), 2231-2250, 2006. doi:10.2514/1.18144
2
E. Gabriel, M. Resch, T. Beisel, R. Keller, "Distributed Computing in a Heterogeneous Computing Environment", in Recent Advances in Parallel Virtual Machine and Message Passing Interface, 494, 1998. doi:10.1007/BFb0056574

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 £125 +P&P)