Parallelization of a dot product with processes (concurrent.futures)

Uses processes to parallelize a dot product is not a very solution because processes do not share memory, they need to exchange data. This parallelisation is efficient if the ratio exchanged data / computation time is low. This example uses concurrent.futures. The cost of creating new processes is also significant.

import numpy
from tqdm import tqdm
from pandas import DataFrame
import matplotlib.pyplot as plt
import concurrent.futures as cf
from teachcompute.ext_test_case import measure_time


def parallel_numpy_dot(va, vb, max_workers=2):
    if max_workers == 2:
        with cf.ThreadPoolExecutor(max_workers=max_workers) as e:
            m = va.shape[0] // 2
            f1 = e.submit(numpy.dot, va[:m], vb[:m])
            f2 = e.submit(numpy.dot, va[m:], vb[m:])
            return f1.result() + f2.result()
    elif max_workers == 3:
        with cf.ThreadPoolExecutor(max_workers=max_workers) as e:
            m = va.shape[0] // 3
            m2 = va.shape[0] * 2 // 3
            f1 = e.submit(numpy.dot, va[:m], vb[:m])
            f2 = e.submit(numpy.dot, va[m:m2], vb[m:m2])
            f3 = e.submit(numpy.dot, va[m2:], vb[m2:])
            return f1.result() + f2.result() + f3.result()
    else:
        raise NotImplementedError()

We check that it returns the same values.

va = numpy.random.randn(100).astype(numpy.float64)
vb = numpy.random.randn(100).astype(numpy.float64)
print(parallel_numpy_dot(va, vb), numpy.dot(va, vb))

-11.963328675229187 -11.963328675229185

Let’s benchmark.

res = []
for n in tqdm([100000, 1000000, 10000000, 100000000]):
    va = numpy.random.randn(n).astype(numpy.float64)
    vb = numpy.random.randn(n).astype(numpy.float64)

    m1 = measure_time("dot(va, vb, 2)", dict(va=va, vb=vb, dot=parallel_numpy_dot))
    m2 = measure_time("dot(va, vb)", dict(va=va, vb=vb, dot=numpy.dot))
    res.append({"N": n, "numpy.dot": m2["average"], "futures": m1["average"]})

df = DataFrame(res).set_index("N")
print(df)
df.plot(logy=True, logx=True)
plt.title("Parallel / numpy dot")

  0%|          | 0/4 [00:00<?, ?it/s]
 25%|██▌       | 1/4 [00:16<00:48, 16.19s/it]
 50%|█████     | 2/4 [00:23<00:21, 10.87s/it]
 75%|███████▌  | 3/4 [00:33<00:10, 10.69s/it]
100%|██████████| 4/4 [01:16<00:00, 23.16s/it]
100%|██████████| 4/4 [01:16<00:00, 19.02s/it]
           numpy.dot   futures
N
100000      0.002819  0.029472
1000000     0.002338  0.011743
10000000    0.007796  0.012140
100000000   0.035683  0.038637

Text(0.5, 1.0, 'Parallel / numpy dot')

The parallelisation is inefficient unless the vectors are big.