Dask and Numba - How to use map partitions efficiently?

I'm trying to speed up my code and improve my understanding of Dask and Numba, I did try to use both on some example I created yet there is no improvement and I can't understand why.

I must say that I'm on on a laptop with four cores so the improvement might not be large but it should be there nonetheless.

More precisely on a Windows 10 laptop, using Python 3.7 and having Numba and Dask in a conda environment.

Here is my code :

import numpy as np
import pandas as pd
from numba import jit
import dask.dataframe as dd

data = np.random.randint(-10, 10, (10**8, 3))
df = pd.DataFrame(data=data, columns=["A", "B", "C"], index=None)
df["F"] = np.random.choice(["apple", "banana", "orange", 
                            "pear","grape","lime","citrus","peach"],10**8)

As you can see it's quite a big dataframe memory wise, this was my way to check if Dask represented on improvement or not. On smaller dataframes (< 200MB) it didn't seem to make a difference at all.

ddf = dd.from_pandas(df,npartitions=12)

@jit
def remove_special_char_with_numba(x):
    return x.replace('r','')

This is a toy example where I try to remove a string from a specific column, Numba does speed up the code compared to vanilla Pandas but doesn't support strings so I can't modify replace or use the nopython mode. Now :

%%timeit
remove_special_char_with_numba(df["F"])

output :

58.9 s ± 9.51 s per loop (mean ± std. dev. of 7 runs, 1 loop each)

Next, my understanding of the following is that Dask separated my dataframe into different chunks/partitions and that it will apply the function on each separated chunk independently. With four cores, it should speed up the process as far as I understand it.

%%timeit
ddf["F"].map_partitions(remove_special_char_with_numba).compute()

output :

45.9 s ± 10.5 s per loop (mean ± std. dev. of 7 runs, 1 loop each)

Now I don't want to be greedy but shouldn't the improvement be larger than that? Am I doing something wrong ?

Thanks