Rpy2, DiceKriging, multiprocessing : memory leak

Question

In order to use the R package DiceKriging from Python with multiprocessing, I wrote something like the following example.

Multiprocessing allows for a significant gain in time, but has the unexpected side effect of quickly increasing memory consumption. When setting flag_pool = False in the following code (hence not using multiprocessing) the memory is stable.

I tried using both Python and R garbage collectors (as suggested here, here, and there) without success.

How could this be avoided?

MWE:

from multiprocessing import Pool

import numpy as np

from rpy2 import robjects
robjects.r("library('DiceKriging')")


# Generate data
def model_to_emulate(X):
    return X[:, 0]**3 - X[:, 1]**2

X = np.random.random_sample((1000, 2))
z = model_to_emulate(X)

list_arg = [[z, X]] * N_gp

# Emulation
flag_pool = True

N_gp = 3
N_repetition = 3

def worker_km(response_design):
    response = response_design[0]
    robjects.globalenv["response"] = robjects.FloatVector(response)

    design = response_design[1]
    df = robjects.r["data.frame"]([robjects.FloatVector(column) for column in
                                   design.T])

    df.names = ["x%d" % ii for ii in xrange(design.shape[1])]
    robjects.globalenv["design"] = df

    return robjects.r("fit = km(design=design, response=response,"
                         "covtype='matern5_2')")

for _ in xrange(N_repetition):
    if flag_pool:
        print "==================== Using Pool."
        pool = Pool(N_gp)
        out = pool.map(worker_km, list_arg)
        pool.close()
        pool.join()
    else:
        print ">>>>>>>>>>>>>>>>>>>> Not using Pool."
        for response_design in list_arg:
            out = worker_km(response_design)
#

Edit: I use ubuntu 12.04.4, python 2.7.3, R 2.14.1 and rpy2 2.2.5.

I asked a similar question here with an MWE that should be easier to run.