FieldView Parallel Performance Information

This image of a horizontal axis wind turbine simulation of the classic National Renewable Energy Laboratory (NREL) reference machine.

Post-processing is a compute intensive task that performs analysis throughout a simulation volume. It is very different from explicit solver methods is not amenable to domain decomposition approaches. Still, the parallel scaling delivered by FieldView Parallel offers 77% time savings when performing typical post-processing tasks with 8 processors in our arbitrary polyhedral element benchmark. For our tetrahedral element model benchmark, 8 processor delivers a 63% time reduction. Both show scale factors exceeding 4x on 8 processors.

As can be seen in the charts provided, scaling is high for reading data and creating surfaces while sweeping operations scale less well (50% scaling). The benchmark has been designed to be representative of common user workflows.

Achieving maximum parallel performance with your simulations:

• Create multi-grid data where the workload is reasonably well distributed or
• Create partitioned solution files using export utilities in select solvers
• Ensure that your client-server setup is working well prior to starting FieldView parallel (note: Xservers such as eXceed or Xwindows are not used for FieldView client-server operation).

Test case benchmark performance:
Tests were performed with tetrahedral data and polyhedral data with two models sizes. Tetrahedral data was tested using solutions from FLUENT (face based results) and AcuSim while FLUENT arbitrary polyhedral data was seperately tested. The sample case consists of airflow through a chamber.

Our chamber is shown below with a zoom in on the grid, shown as a crinkle surface in FieldView.

The benchmark example puts FieldView through the following tasks:

1. Read in the data
2. Create surfaces
3. Create coordinates
4. Sweep surfaces

The images below show the time savings from using FieldView Parallel and the parallel efficiency demonstrated in these cases.

Tetrahedral Model


View Larger

Polyhedral Model


View Larger

As can be seen, parallel performance is greatest where communication between processors is minimized. So operations such as data reading, creation of surfaces, creating iso-surfaces, and function calculations will scale very well. Operations that will scales less well include the generation of streamlines and particle paths, feature extractions such as the identification of vortex cores, surface flows and separation/reattachment lines. Also, the creation of boundary and computational surfaces and sweeping operations also require significant inter-processor communication by their very nature. The ability to save large amounts of time using FieldView Parallel provides great productivity improvement for organizations facing post-processing related workflow bottlenecks due to the size and volume of their CFD simulation data.