TUFLOW HPC Advice

A common question we receive is, "how can I speed up my TUFLOW HPC simulation?" This Insights article lists and demonstrates some general tips how to optimise a TUFLOW HPC model (run using GPU hardware) to achieve a faster simulation time.

1. Review the domain extent and if necessary, update the domain orientation and extents. Using a domain that is compactly nested around the extent of your active model area (Code == 1) will result in the best allocation of compute load to the GPU card CUDA cores responsible for the simulation.
2. Check the simulation minimum dt output result to assess whether there are input errors in your model that are artificially forcing TUFLOW HPC to reduce to a lower than necessary timestep in order to maintain a stable solution.
3. If you are using Sub-Grid Sampling (SGS), check your SGS frequency is not excessive. By default at the time of writing, TUFLOW HPC will adopt a SGS Sample Target Distance that matches the minimum grid resolution of your input DEM. This value can be increased if the resulting SGS Sample Frequency is excessively large or if the physical geometry in your model is mild enough to not warrant such high frequency sampling. Parameter sensitivity testing is a useful exercise to identify the maximum SGS Sample Target Distance, or conversely minimum SGS Sample Frequency, that achieves a converged solution.
4. Check your model 2D cell size is the largest value that achieves a converged solution at your area of interest. If you are not familiar with cell size convergence testing, we recommend watching our Cell Size Optimisation Webinar recording. 
5. Review the GPU card you are using for your simulation. The TUFLOW community have collectively shared GPU card speed benchmark results with us in order to create a GPU card performance database. Find your GPU card in the TUFLOW Wiki Hardware Benchmarking results table (ordered fastest to slowest, top to bottom). If your GPU card is not already included in the list, participate in the benchmarking and email us your results so we can add your information to the database.    
6. Can multiple simulations be run on a single GPU card? It is a common misconception that a single GPU card can only accommodate a single TUFLOW HPC simulation. Depending on the GPU RAM demand of your model, relative to the GPU RAM capacity of your card, multiple models can be run on a single card. When running two simulations on a single GPU card, each model simulation will run slower, though collectively from a project perspective the complete suite of simulations will likely finish quicker than if the model simulations were configured to run as a single model execution in series.

   The video below discusses each of the above topics in detail and also demonstrates their impact on simulation time. Please watch it to learn more. 

   The above six model design optimisation and hardware execution tips are common simple things that will improve project execution efficiency when using TUFLOW HPC. This is however just the tip of the iceberg. There are many more model specific ways to design a TUFLOW HPC model that is optimisted for model speed, computer memory (CPU and GPU) and also the minimum data footprint. If you would like to learn more, we highly recommend attending one of our instructor led training courses to speak with and learn from one of our modelling experts. Please see our Training Catalogue for a list of our available courses and/or email training@tuflow.com.