Project description
The research project HICFD is funded by the German Ministry for Education and Research within the programme "IKT 2020 - Research and Innovation". The project´s objective is to develop new methods and tools for the analysis and optimization of the performance of parallel CFD codes on high performance computer systems with many-core processors (HPC many-core architectures) and to exemplarily apply these to DLR´s CFD programs TAU (computation of external flows) and TRACE (simulation of internal flows).
In the workpackages of the project it is examined how the performance of parallel CFD codes can be increased by the optimal exploitation of all parallelism levels. On the highest, with MPI (Message Passing Interface) parallelized level an intelligent mesh partitioning in order to improve the load balance between the MPI processes is promising. For the block-structured grids used in TRACE a many-core compatible partitioning tool is developed.
Furthermore, on the level of the many-core architecture, highly scaling, hybrid OpenMP/MPI methods (OpenMP: Open Multi-Processing, standard for shared-memory programming) are implemented for the CFD solvers TAU and TRACE. Within the block-structured CFD code TRACE the iterative algorithm for the solution of linear systems of equations has to be optimized, among other things by preconditioning methods adequate for many-core architectures. For the unstructured CFD solver TAU it is investigated if the parallel efficiency of the multigrid algorithm used can be increased.
On the level of the processor cores a pre-processor is developed which makes a comfortable exploitation of the parallel SIMD units (Single Instruction Multiple Data) possible, also for complex applications. For a detailed performance examination of SIMD operations the tracing abilities of the performance analysis suite Vampir is further developed.
Beyond the specific project goals a best-practice report is composed which treats the topics "hybrid parallelization" and "SIMD exploitation". This document sums up the experiences from the HICFD project and serves as guidance how simulation codes also from other disciplines than CFD can be optimized for efficient execution on many-core architectures.
Project partners
Project duration
01.01.2009 - 31.12.2011