Initialization of FlorisModel is slow for large farms

[Bartdoekemeijer]

I'm currently upgrading my internal workflows to FLORIS v4. The new features are awesome. One thing I'm struggling with is the initialization time of FlorisModel objects for large farms.

Here's a simple test case:

from time import perf_counter as timerpc
from matplotlib import pyplot as plt
import numpy as np
from floris import FlorisModel, WindRose


# Set-up a common FLORIS model
fmodel = FlorisModel("inputs/gch.yaml")

# Decide on how farm sizes that we want to evaluate
n_turbines_array = [3, 10, 20, 50, 200, 500, 1000]

# Set up a wind rose
wd_array = np.arange(0.0, 360.0, 3.0)
ws_array = np.arange(1.0, 25.01, 1.0)
wd_grid, ws_grid = np.meshgrid(wd_array, ws_array, indexing="ij")
freq_grid = np.ones_like(wd_grid)
wind_rose = WindRose(wind_directions=wd_array, wind_speeds=ws_array, ti_table=0.06, freq_table=freq_grid)

# Compute load time for different layout sizes
timings = np.zeros_like(n_turbines_array)
for nii, n_turbines in enumerate(n_turbines_array):
  fmodel.set(
    layout_x=600*np.arange(n_turbines),
    layout_y=np.zeros(n_turbines),
    wind_data=wind_rose,
  )
  fm_dict = fmodel.core.as_dict()

  t0 = timerpc()
  FlorisModel(fm_dict)
  timings[nii] = timerpc() - t0
  print(f"Timing for {n_turbines} turbines: {timings[nii]:.1f} s.")

# Plot timings
fig, ax = plt.subplots()
ax.plot(n_turbines_array, timings, '--o')
ax.set_xlabel("Number of turbines (-)")
ax.set_ylabel("Model initialization time (s)")
ax.set_title(f"FlorisModel initialization time with n_findex={wind_rose.n_findex}")
ax.grid(True)
plt.show()

For parallel computing, I may split up the task into, e.g., 120 FlorisModel objects that each process a different set of findex entries. However, when modeling 500+ turbines (think parts of the Dutch North Sea), it may take a minute to load one FlorisModel. Parallelizing this to 120 tasks then adds 2 CPU hours of computation time overhead.

I found that the initialization time is not sensitive to n_findex. There is a linear correlation in the number of turbines and the initialization time.

Is there a way to speed up model initialization?

[misi9170]

Hi @Bartdoekemeijer , thanks for this! This is not something I'd been aware of. Running your code on a Mac (and python 3.10.14), I'm getting surprisingly different results:

The overall speed difference could be down to just whatever else is running in the background, but I'm surprised that the insensitivity to n_findex isn't the same.

Still, I'll look into profiling the code to see where the time is spent on initialization to see if there are easy opportunities for speed up

[misi9170]

Here's the full script I'm running, based on yours:

from time import perf_counter as timerpc
from matplotlib import pyplot as plt
import numpy as np
from floris import FlorisModel, WindRose


# Set-up a common FLORIS model
fmodel = FlorisModel("inputs/gch.yaml")

# Decide on how farm sizes that we want to evaluate
n_turbines_array = [3, 10, 20, 50, 200, 500, 1000]

# Set up a wind rose
wd_array = np.arange(0.0, 360.0, 3.0) # 3.0 degrees will produce n_findex=3000
ws_array = np.arange(1.0, 25.01, 1.0)
wd_grid, ws_grid = np.meshgrid(wd_array, ws_array, indexing="ij")
freq_grid = np.ones_like(wd_grid)
wind_rose = WindRose(wind_directions=wd_array, wind_speeds=ws_array, ti_table=0.06, freq_table=freq_grid)

# Compute load time for different layout sizes
timings = np.zeros_like(n_turbines_array)
for nii, n_turbines in enumerate(n_turbines_array):
  fmodel.set(
    layout_x=600*np.arange(n_turbines),
    layout_y=np.zeros(n_turbines),
    wind_data=wind_rose,
  )
  fm_dict = fmodel.core.as_dict()

  t0 = timerpc()
  FlorisModel(fm_dict)
  timings[nii] = timerpc() - t0
  print(f"Timing for {n_turbines} turbines: {timings[nii]:.1f} s.")

# Plot timings
fig, ax = plt.subplots()
ax.plot(n_turbines_array, timings, '--o', label=f"n_findex={wind_rose.n_findex}")
ax.set_xlabel("Number of turbines (-)")
ax.set_ylabel("Model initialization time (s)")
#ax.set_title(f"FlorisModel initialization time with n_findex={wind_rose.n_findex}")

# Reinstantiate the FlorisModel
fmodel = FlorisModel("../inputs/gch.yaml")

# Decide on how farm sizes that we want to evaluate
n_turbines_array = [3, 10, 20, 50, 200, 500, 1000]

# Set up a wind rose
wd_array = np.arange(0.0, 360.0, 30.0) # 30.0 degrees will produce n_findex=300
ws_array = np.arange(1.0, 25.01, 1.0)
wd_grid, ws_grid = np.meshgrid(wd_array, ws_array, indexing="ij")
freq_grid = np.ones_like(wd_grid)
wind_rose = WindRose(wind_directions=wd_array, wind_speeds=ws_array, ti_table=0.06, freq_table=freq_grid)

# Compute load time for different layout sizes
timings = np.zeros_like(n_turbines_array)
for nii, n_turbines in enumerate(n_turbines_array):
  fmodel.set(
    layout_x=600*np.arange(n_turbines),
    layout_y=np.zeros(n_turbines),
    wind_data=wind_rose,
  )
  fm_dict = fmodel.core.as_dict()

  t0 = timerpc()
  FlorisModel(fm_dict)
  timings[nii] = timerpc() - t0
  print(f"Timing for {n_turbines} turbines: {timings[nii]:.1f} s.")

ax.plot(n_turbines_array, timings, '--o', label=f"n_findex={wind_rose.n_findex}")
ax.set_xlabel("Number of turbines (-)")
ax.set_ylabel("Model initialization time (s)")

ax.grid(True)
ax.legend()
plt.show()

[paulf81]

Just wanted to add my result using Python 3.11.9 (main, Apr 19 2024, 11:44:45) [Clang 14.0.6 ] on darwin

[misi9170]

Also tagging #885 as it's likely related

[misi9170]

I've implemented an improvement to address this in #966

[Bartdoekemeijer]

Solved by #966, thank you Misha! Closing this discussion.

[misi9170]

No problem, thanks for flagging the issue!