23.8 Parallel Processing for the Discrete Phase Model

ANSYS FLUENT offers two modes of parallel processing for the discrete phase model: the Shared Memory and the Message Passing options under the Parallel tab, in the Discrete Phase Model dialog box. The Shared Memory method is suitable for computations where the machine running the ANSYS FLUENT host process is an adequately large, shared memory, multiprocessor machine. The Message Passing option is enabled by default and is suitable for generic distributed memory cluster computing.

When tracking particles in parallel, the DPM model cannot be used with any of the multiphase flow models (VOF, mixture, or Eulerian) if the Shared Memory option is enabled. (Note that using the Message Passing option, when running in parallel, enables the compatibility of all multiphase flow models with the DPM model.)

The Shared Memory option (Figure  23.8.1) is implemented using POSIX Threads ( pthreads) based on a shared memory model. Once the Shared Memory option is enabled, you can then select along with it the Workpile Algorithm and specify the Number of Threads. By default, the Number of Threads is equal to the number of compute nodes specified for the parallel computation. You can modify this value based on the computational requirements of the particle calculations. If, for example, the particle calculations require more computation than the flow calculation, you can increase the Number of Threads (up to the number of available processors) to improve performance. When using the Shared Memory option, the particle calculations are entirely managed by the ANSYS FLUENT host process. You must make sure that the machine executing the host process has enough memory to accommodate the entire mesh.

Note that the Shared Memory option is not available for Windows 2000.

Please note that the Workpile Algorithm option is not available with the wall film boundary condition. It will be disabled automatically when chosing to simulate a wall film on a wall.

Figure 23.8.1: The Shared Memory Option with Workpile Algorithm Enabled

The Message Passing option enables cluster computing and also works on shared memory machines. With this option enabled, the compute node processes perform the particle work on their local partitions. Particle migration to other compute nodes is implemented using message passing primitives. There are no special requirements for the host machine. Note that this model is not available if the Cloud Model option is turned on under the Turbulent Dispersion tab of the Set Injection Properties dialog box. When running ANSYS FLUENT in parallel, by default, pathline displays are computed in serial on the host node. Pathline displays may be computed in parallel on distributed memory systems if the Message Passing parallel option is selected in the Discrete Phase Model dialog box.

You may seamlessly switch between the Shared Memory option and the Message Passing option at any time during the ANSYS FLUENT session.

In addition to performing general parallel processing of the Discrete Phase Model, you have the option of implementing DPM-specific user-defined functions in parallel ANSYS FLUENT. For more information about the parallelization of DPM UDFs, see this section in the separate UDF Manual.

When using the Message Passing option you can make use of ANSYS FLUENT's automated load balancing capability by giving an appropriate weight to the particle steps in each cell. In this case the number of particle steps in each partition is considered in the load balancing procedure. Further details can be found in Section  32.5.5.