### TurbSim CPU Times

Posted:

**Thu Dec 23, 2010 1:24 pm**I occasionally get questions regarding the approximate time that it takes to create large grids in TurbSim, so I thought I would post some plots to illustrate the relationship between the number of grid points (on an n by n grid) and computation time. The computation time is dominated by the Cholesky decomposition of an n^2 by n^2 matrix at each time step.

These plots were created using the IECKAI model with TurbSim v1.50 (non-IEC models can take at least 2 times longer because they compute a Cholesky decomposition on the v- and w- wind components as well). I used AnalysisTime = 600 seconds and TimeStep = 0.05 seconds, and I ran this on a 32-bit WinXP Duo Core CPU (2.83 GHz) with 4 GB RAM. For each simulation, I used the same input file, changing only NumGrid_z and NumGrid_y (which are equal in each case), and plotted the CPU time listed at the end of the simulation.

Here are the results, plotted in two different ways:

The 41 x 41 grid required a little less than 0.5 GB RAM, so all of these cases ran well within memory.

NOTE: I used a 10 m/s mean hub-height wind speed and 140 m x 140 m grids for these simulations. Interestingly, when I ran the 17 x 17 case using 12.368 m/s and a 50 m x 50 m grid instead (everything else was the same), the CPU time increased from approximately 39 s to 94 s. This is due to the fact that the coherence matrix is more difficult to factor. Keep this in mind when comparing specific CPU times with these plots.

These plots were created using the IECKAI model with TurbSim v1.50 (non-IEC models can take at least 2 times longer because they compute a Cholesky decomposition on the v- and w- wind components as well). I used AnalysisTime = 600 seconds and TimeStep = 0.05 seconds, and I ran this on a 32-bit WinXP Duo Core CPU (2.83 GHz) with 4 GB RAM. For each simulation, I used the same input file, changing only NumGrid_z and NumGrid_y (which are equal in each case), and plotted the CPU time listed at the end of the simulation.

Here are the results, plotted in two different ways:

The 41 x 41 grid required a little less than 0.5 GB RAM, so all of these cases ran well within memory.

NOTE: I used a 10 m/s mean hub-height wind speed and 140 m x 140 m grids for these simulations. Interestingly, when I ran the 17 x 17 case using 12.368 m/s and a 50 m x 50 m grid instead (everything else was the same), the CPU time increased from approximately 39 s to 94 s. This is due to the fact that the coherence matrix is more difficult to factor. Keep this in mind when comparing specific CPU times with these plots.