Annual Report

2.4  Software engineering

• Source management: AVBP is developed jointly at Institut Francais du Pétrole and CERFACS but is used at many other institutions in Europe. Source management is performed only at CERFACS: typically two versions are produced and tagged every year using CVS. Regular meetings take place between IFP and CERFACS to define the evolution of versions.
• CTEA are elementary test cases (convection, diffusion, time advancement) which are run on a weekly basis automatically at night and compare versions during their modifications.
• Quality Program Forms (QPF): QPFs are more complete tests of the code which are performed every six months to ensure non regression of versions. QPFs include laminar flames, tubes, homogeneous turbulence, shear layers, acoustic waves, Karmann street, etc.
• Web user page: all documentation related to AVBP is summarized on a specific web page which can be accessed by all users and developers. The web site contains the handbooks, a page to report problems, a user's guide, various tutorials, all QPFs of all versions, reports and PhDs directly linked to AVBP, etc.
• Optimization and parallelization: most present platforms for high-performance computing use parallel codes and AVBP follows the same path. This puts constraints on the algorithms used for LES. For example, in the field of two-phase flows, using Euler Euler formulations instead of Euler Lagrange is typically dictated partly by computer architecture considerations (Section 2.2).
• Visualization: LES generates large data fields but conversely to RANS, LES generates many such fields to retain the unsteady character of the flow. A typical LES can create 200 snapshots on a 2 million points grid. Transporting such fields and storing them becomes difficult. CERFACS has introduced different solutions for this: (1) special tools have been created to reduce the size of data fields (by using coarser grids) in order to transport them back to the user and visualizing the results in a degraded but faster mode (2) experimentations performed with CINES on remote visualization with OpenGL Vizserver: data remains at CINES and is processed on a specific SGI machine and only pictures are transferred in real time to the user. The hardware and software needed for this as well as the results obtained for AVBP data are described in [1] (see also 2.5 of the Computer Support Group chapter). They have also been presented in the specialized press by SGI.
• More generally CERFACS continues to collaborate intensively with the French national super computing centers CINES and IDRIS but also with CEA to develop high performance computing for reacting flows on parallel computers. AVBP or NTMIX are used to benchmark machines for CINES and IDRIS for example. Regular meetings offer opportunities to CERFACS users to discuss the evolutions of the computing centers. This collaboration also extends to the USA: in 2002, Pr W.C. Reynolds, head of the ASCI project at Stanford (which is devoted to computations of combustion in gas turbines on massively parallel machines) came to CERFACS to present the results of the project. In 2003, a very special program took place between CINES and CERFACS: in this 'extreme combustion' project, very large LES were performed (up to 8 million points) to demonstrate the present limits of existing platforms. Interestingly, results have matched what CERFACS scientists have already observed at Stanford in 2002 during the summer program: the present limits have been found to be essentially at the preprocessing and postprocessing levels. Generating very large hybrid grids (more than 8 million points) and partitioning them for parallel computers for example is a key problem at the present time.