Why is it faster to first pickle data before sending a dict than sending a dict directly in Python 3?

[Martin Bammer]

Following test program:
import time
import marshal
try:
    import cPickle as pickle
    str_ = unicode
except:
    import pickle
    str_ = str

def TestPrc(rd):
    d = rd.recv()
    print("OK")

def TestPrc2(rd):
    d = pickle.loads(rd.recv())
    print("OK")

if __name__ == "__main__":
    from multiprocessing import freeze_support, Process, Pipe
    freeze_support()
    d = { str_(x) : u"Hello World" + str_(x + 1) for x in range(1000000) }
    wr, rd = Pipe()
    p = Process(target = TestPrc, args = ( rd, ))
    p.start()
    t1 = time.time()
    wr.send(d)
    print("#1", time.time() - t1)
    p.join()
    print("#2", time.time() - t1)
    p = Process(target = TestPrc2, args = ( rd, ))
    p.start()
    t1 = time.time()
    wr.send(pickle.dumps(d, pickle.HIGHEST_PROTOCOL))
    print("#3", time.time() - t1)
    p.join()
    print("#4", time.time() - t1)

I get the following results:
Python 2.7:
('#1', 0.33500003814697266)
OK
('#2', 0.7890000343322754)
('#3', 0.36300015449523926)
OK
('#4', 0.8059999942779541)

Python 3.4:
#1 0.7770781517028809
OK
#2 1.4451451301574707
#3 0.7410738468170166
OK
#4 1.3691368103027344

Python 3.6:
#1 0.681999921798706
OK
#2 1.1500000953674316
#3 0.6549999713897705
OK
#4 1.1089999675750732

The results show that in Python 3 it is faster to do the (un-)pickling in the Python code. I would expect to have no real performance difference, or at least the more straight forward way to send Python objects directly to be a bit faster, as it is in Python 2.
Some other interesting results from this example:
- Python 2 is much faster
- At least Python 3.6 is much faster than Python 3.4

Regards,

Martin