Added demo code for parallelizing the data generation code.

src/joblib_demo.py

@@ -0,0 +1,69 @@
+#!/usr/bin/env python
+
+"""This shows how the data generation can be parallelized. For 16 runs, I'm
+getting twice the speed up as compared to the sequential method."""
+
+from multiprocessing import Pool
+
+import numpy as np
+from joblib import Parallel, delayed
+
+from model import QuietStandingModel
+from measured_data import DataGenerator
+from utils import timeit
+
+num_nodes = 51
+duration = 5.0
+
+ref_noise_std = 0.03
+platform_accel_noise_std = 0.0
+coordinate_noise_std = 0.0
+speed_noise_std = 0.0
+torque_noise_std = 0.0
+
+# Generate the symbolic dynamic model.
+model = QuietStandingModel()
+model.derive()
+
+num = 12
+
+
+def gen(platform_pos_mag):
+
+    # This involves generating the rhs function and is slower.
+    data = DataGenerator(duration, num_nodes, ref_noise_std,
+                         platform_pos_mag, model=model)
+    # This step is fast.
+    data.generate(platform_accel_noise_std, coordinate_noise_std,
+                  speed_noise_std, torque_noise_std)
+
+    # NOTE : You can't return the data object here because something inside
+    # it can't be pickled, so I return the dictionary instead.
+    return data.measured
+
+platform_pos_mags = np.linspace(0.0, 0.3, num=num)
+
+pool = Pool()
+
+
+@timeit
+def run_joblib():
+    return Parallel(n_jobs=-1)(delayed(gen)(platform_pos_mag) for
+                               platform_pos_mag in platform_pos_mags)
+
+
+@timeit
+def run_multiprocessing():
+    return pool.map(gen, [p for p in platform_pos_mags])
+
+
+@timeit
+def run_sequential():
+    for p in platform_pos_mags:
+        gen(p)
+
+
+if __name__ == "__main__":
+    run_sequential()
+    run_multiprocessing()
+    run_joblib()