Seeking explanation to Dask vs Numpy vs Pandas benchmark results

Question

I am trying to benchmark the performance of dask vs pandas .

def make_pandas(n):
    df = pd.DataFrame(np.random.randint(10, size=(n, 3)))
    return df

def make_dask(n):
    df = da.from_array(np.random.randint(10, size=(n, 3)), chunks=10)
    return df

def make_numpy(n):
    return np.random.randint(10, size=(n, 3))

def sum_pandas(x): return x[0].sum()
def sum_dask(x): return x[1].sum()
def sum_numpy(x): return x[2].sum()

perfplot.show(
    setup=lambda n: [make_pandas(n), make_dask(n), make_numpy(n)],
    kernels=[sum_pandas, sum_dask, sum_numpy],
    n_range=[2**k for k in range(2, 15)],
    equality_check=False,
    xlabel='len(df)')

Can someone explain the results:

Increasing the chunks to 1000, 8000 and 10000 gives these respectively:

• Processor: Intel® Core™ i5-7300HQ CPU @ 2.50GHz × 4
• Memory: 7.7 GiB
• Python: 3.5.2
• pandas: 0.21.0
• numpy: 1.13.1
• dask: 0.19.0

Isn't dask supposed to parallelize and be better as size increases?

Answer 1

chunks关键字是chunksize的缩写，不是chunk的数量