python - python joblib & random walk - a performance of [CONCURRENT]-process scheduling

Question

Here is my python-3.6 code for simulating a 1D reflected random walk, using the joblib module to generate 400 realizations concurrently across K workers on a Linux cluster machine.

I note, however, that the runtime for K=3 is worse than for K=1, and that the runtime for K=5 is even worse!

Can anyone please see a way to improve my use of joblib?

from math import sqrt
import numpy as np
import joblib as jl
import os

K = int(os.environ['SLURM_CPUS_PER_TASK'])

def f(j):
    N = 10**6
    p = 1/3
    np.random.seed(None)
    X = 2*np.random.binomial(1,p,N)-1   # X = 1 with probability p
    s = 0                               # X =-1 with probability 1-p
    m = 0 
    for t in range(0,N):
        s = max(0,s+X[t])
        m = max(m,s)
    return m

pool = jl.Parallel(n_jobs=K)
W = np.asarray(pool(jl.delayed(f)(j) for j in range(0,400)))
W