The project work included the development and testing of reliable Linux scripts for efficient long-term usage: One objective was the efficient and error-free adjustment of the Cray batch scripts and OpenFOAM 's dictionaries to, for example, varying core numbers and different use cases; another challenge was how to extract the relevant pieces of information from the Cray Profiler output files in order to easily import them into Excel.
The main topics were parallel mesh generation see section 2 , domain decomposition and mesh reconstruction section 3 and turbulence modelling section 4. Furthermore, we report on our experiences and lessons learned during this SHAPE pilot project in section Parallel mesh generation with snappyhexmesh The parallel tool snappyhexmesh  generates meshes snapped to surfaces defined, for example, by CAD data in STL format.
It offers the possibility to vary the number of refinement levels r of the initial block mesh as well as the number of cells between refinement levels c and the number of added surface layers s. We were interested in how these three parameters influence the final number of mesh cells, the wall time and the memory used by snappyhexmesh. We studied a very simple geometry, a cube see Figure 2 and subsection 2. Our goal was to find an approximate formula fitting the number of mesh cells n r.
Therefore, we make the following approach for the total number n of mesh cells depending on the number of refinement levels r: n r r A V A z z z Figure 3: Fitting n r. The horizontal axis is the refinement level r, and the vertical axis is the cell number n r. The mock-up Mock-up models are typically physical constructions made of wallboard with reduced details.
Main purpose of these models is, to encourage people to test handling with the wooden devices of the later clean room in an early stage of design and development. Figure 4 presents the complex geometry of a filling line in a clean room. Filling devices, hoppers, sorter devices and transport chain are displayed.
Right: detail. Figure 5 shows an automatically generated mesh by the OpenFOAM tool snappyhexmesh with refinement level 2 and with 4 surface layers see detail image on all boundaries. Mesh quality and surface layer coverage is quite satisfying, just some small details have some missing surface layers. Domain decomposition and mesh reconstruction The serial tools decomposepar, reconstructparmesh and reconstructpar are commonly used for mesh decomposition after running snappyhexmesh and for reconstruction the mesh and the fields after the solver run.
They are major bottlenecks: We found insufficient performance of these serial tools on large meshes with more than 40 million 40 M cells. Furthermore, they require more memory than usually available: decomposepar runs out of memory on an ordinary 64 GB node when decomposing snapped meshes with more than 40 M cells. Similarly, reconstructparmesh can handle meshes only up to 30 to 40 M cells on a 64 GB node. In addition, 4. A possible workaround of the memory problem would have been the utilization of special pre- and postprocessing nodes with 1 TB memory available on Hermit, but we found another solution and successfully applied it: The usage of the parallel OpenFOAM tools renumbermesh and potentialfoam.
This means that the serial processes can partly be substituted by parallel processes. Turbulence modelling and scaling tests We took a nearly direct numerical simulation DNS of the Karman vortex street as benchmark for the turbulence model see Figure 6. The Karman vortex street seems to be the most suitable case for transient flow simulation benchmarks because it is well investigated, and a simple formula for the vortex frequency is also well-known see below. DNS is the nearest approximation to reality.
Figure 6: Karman vortex street. The Karman simulation shows good compliance with the theory of the Karman vortex street. As we did not utilize wall functions to model the flow near walls, we needed a finer mesh resolution with surface layers near the walls. OpenFOAM s tool yplusles gives values between 0.
Left: flow tubes. Right: pressure iso-surfaces. Left: streamlines and k- isosurfaces. Right: detail with streamlines A stationary adaption of air flow in clean rooms is shown in Figure 8. The simulation was run with the simplefoam k-omega-sst solver. Detached eddies and other typical transient effects can t be resolved by this stationary solutions.
Another issue are the cellsizes of the nearest cells to the surface surface contact. For this CFD case, wall-function surface cells should have 30mm thickness! So it is impossible to resolve CAD structures below 60mm.
Buy CFD-Modellierung: Grundlagen und Anwendungen bei Strömungsprozessen (German Edition) on tworchutztode.gq ✓ FREE SHIPPING on qualified orders. Editorial Reviews. From the Back Cover. In diesem kompakten Lehrbuch wird die Methodik der note taking and highlighting while reading CFD-Modellierung: Grundlagen und Anwendungen bei Strömungsprozessen (German Edition).
The Reynolds Number Re is , after all. This seems to be too small for RANS models. Good strong scaling only could be performed up to cores e. The reasons for the poor performance 6. Initially there was an issue where the computing time ran out and processes ended without any error messages of the profiler or the queue or Tier-0 system executables. No results were written on the file system, but wall-clock time ran out. This was a source of some frustration to the team, however after some investigation it was discovered that the Tier-0 system s disk quota was being exceeded due to the number of files being generated by OpenFOAM.
Every process generates about 50 files dependent on the number of written time steps so that massively parallel jobs with cores generate about files. This is the reason why only a small number of scaling tests could be performed. Only some snappyhexmesh scaling jobs ran without quota problems.
The results can be seen in following tables. Table 2: Strong Scaling Tests Number of cores Number of cells Wall-clock time Speed-up vs the first one Number of nodes Number of processes Speed-up ideal Speed-up efficiency fault fault fault Table 3: Scaling Tests with constant core number for estimation of the maximum possible cell number Number of cores Number of cells Wall clock time [s] Speed-up vs the first one Number of nodes Number of process Speed-up efficiency Approx.
There were frequent telephone calls, often daily, and weekly or bi-weekly meetings at the HLRS. Prediction of mesh sizes, processor resources and wall-clock time of all OpenFOAM processes helps to optimize the HPC case and to save much money and time. The results will help to select the most appropriate solver for handling air flows in clean rooms at a maximum of accuracy and a minimum of resources. See also section 5.
The industrial partner was quite surprised that the work on a Tier-0 system appeared not as predictable as expected. Furthermore, the queuing time strongly depends on the workload of other users. Another time consuming issue is the occasional overloading of the front nodes, if users run commands that demand more resources than intended.
Therefore, more single processor shared front nodes appear desirable.
Another lesson learned is that the work took more time than foreseen. The industrial partner was so intent upon preparing the OpenFOAM cases for efficient, long-term usage that the time was quite short for this SHAPE project: there were only three months between the start of the preparatory access at the end of January and the writing of the final documents beginning in May. Furthermore, it cannot be assumed that the partners from industry have no other duties besides their SHAPE projects. Altogether, the work has taken much more time than two personal months up to now, and the preparatory access project is still running till the beginning of August.
We adequately integrated the new results and methods in our design processes.
With a better knowledge of the airflow in our clean rooms, we are now able to enhance the design of our filling machines according to our customers' requirements. This optimized set of parameter files and OpenFOAM dictionaries will be used as template for many other clean room geometries and airflow cases. Solver independent findings of this project, like optimized meshing and decomposition, can be implemented to solve new problems and tasks like multiphase studies with OpenFOAM's interfoam solver, for example. In the future, airflow simulations will help to replace most of the expensive smoke studies in production machines at the customer's site.
Simulations will also attend the design process in an early state and support the whole CAE process. Most of the negative developments can be avoided at an early stage, and trial and error cycles and expensive redesigns and reconstructions can be minimized. Conclusions It was a very interesting and challenging project. References     wickie. CastNet: Modelling platform for open source solver technology. June Outline Introduction. Comparison of CFD models for multiphase flow evolution in bridge scour processes A. Valero, F.
Pre-processing in openfoam, mesh generation. Using blockmesh,. Computational fluid dynamics CFD is the science of predicting fluid flow, heat and mass transfer, chemical reactions, and related phenomena by solving numerically.
Discovery Laboratory Multiphase Flow - Appendices 1. Creating a Mesh 1. What is a geometry? The geometry used in a CFD simulation defines the problem domain and boundaries; it is the area 2D or volume. Problem Statement In order to satisfy production and storage requirements, small and medium-scale industrial facilities commonly occupy spaces with ceilings ranging between twenty and thirty feet in height.
Pullinger and J. Computational Fluid Dynamics CFD is the science of predicting fluid flow, heat and mass transfer, chemical reactions,. Kaltenbacher 1, A. Reppenhagen 2,. Raach and S. All rights reserved. CFD: What is it good for? Best practices for efficient HPC performance with large models Dr. D Ahmad Pesaran, Ph.
EVS October -8, 8, Yokohama,. Inlet and outlet boundaries. Der Schwarze schnackselt gerne. Auch Deutschland hatte ja viele Kolonien in Afrika. Wer sich selbst zur Zielscheibe macht, darf sich nicht wundern, wenn andere danach trachten, ins Schwarze zu treffen.