Setting up and launching a parallel simulation has become much simpler and easier in CFD-FASTRAN V2007.  This note discusses some of these developments.

Let’s start with things that have not changed.  There are still two versions of CFD-FASTRAN solvers for parallel cases.  The difference between the two versions is the underlying parallel communication mechanism, the choice of which is decided by the type of mesh used.  Structured mesh models use MDICE (Multi-Disciplinary Computing Environment) and unstructured mesh models use MPI (Message Passing Interface).  MDICE allows different applications to “talk” to each other and hence allows multi-disciplinary simulations such as coupled flow-structure simulations possible without integrating all features (disciplines) into a single code.

The following steps have been removed from V2007:

• Need to run special scripts as part of the installation procedure
• Need to launch MDICE registry and daemons
• Need to launch parallel job only from the command prompt using different scripts such as CFD-FASTRAN-SOLVERP and fastran-mpi

These operations are automatically performed by the script behind the scenes, and hence the user does not have to perform these steps. 

The following options have been added to setup parallel cases for CFD-FASTRAN:

• Can setup both MDICE-based and MPI-based simulations using the improved parallel configuration panel (located on the Run tab)
• Can launch a parallel job (MDICE or MPI) from CFD-FASTRAN-GUI or command prompt
• A single command-prompt command for all (serial and parallel) FASTRAN job submissions (CFD-FASTRAN-SOLVER –dtf <file name>)
• Can easily edit available hosts through CFD-FASTRAN-GUI

Now, we look at some of the settings available in the new Parallel Setup Panel.  This panel can be reached through the Configure Parallel Run option available in the Run tab when parallel is selected as the Run Type.  Depending on the type of mesh used, the MDICE or MPI panel will become available.  The Host List can be edited by clicking on the Edit button on the left hand side of the parallel setup panel.

In MDICE-based parallel simulations, the Distribute Zones button in the Zone Assignment box performs load balancing based on the number of cells.  This just assigns available zones to different nodes.  The load factor column gives an indication of the maximum speed up that can achieved through a particular distribution.  The user can review this and reassign the zones to get as close to a load factor of 1 for all zones.  Therefore, it is advisable that the user plans ahead to avoid the possibility of having large zones that tilts the load factor unfavorably.  Also note that the chimera and stress zones should be assigned to an exclusive processor.  At this time, chimera and stress calculations are performed in serial only.

In MPI-based parallel simulations, domain decomposition is performed using dtf_decompose.  Several options for domain decomposition are available.  The Multilevel Graph Partitioning Scheme provides high quality partitioning with less CPU time and performs work load balancing.  If you would like to know more about domain decomposition, please follow this link. 

Regards,
Abraham Meganathan
ESI CFD Support Team