psf
Benchmark with disabled GIL build and -Xgil=1 isn't enabling GIL
It seems that when you run PYTHON_GIL=1 python benchmark.py the benchmarked functions are being run with the GIL disabled and ignoring the flag.
Benchmark Code
import pyperf
from multiprocessing import Process
from threading import Thread
try:
import _xxsubinterpreters as interpreters
except ImportError:
import _interpreters as interpreters
import itertools
DEFAULT_DIGITS = 2000
icount = itertools.count
islice = itertools.islice
def gen_x():
return map(lambda k: (k, 4 * k + 2, 0, 2 * k + 1), icount(1))
def compose(a, b):
aq, ar, as_, at = a
bq, br, bs, bt = b
return (aq * bq,
aq * br + ar * bt,
as_ * bq + at * bs,
as_ * br + at * bt)
def extract(z, j):
q, r, s, t = z
return (q * j + r) // (s * j + t)
def gen_pi_digits():
z = (1, 0, 0, 1)
x = gen_x()
while 1:
y = extract(z, 3)
while y != extract(z, 4):
z = compose(z, next(x))
y = extract(z, 3)
z = compose((10, -10 * y, 0, 1), z)
yield y
def calc_ndigits(n=DEFAULT_DIGITS):
return list(islice(gen_pi_digits(), n))
test ="""
import itertools
DEFAULT_DIGITS = 2000
icount = itertools.count
islice = itertools.islice
def gen_x():
return map(lambda k: (k, 4 * k + 2, 0, 2 * k + 1), icount(1))
def compose(a, b):
aq, ar, as_, at = a
bq, br, bs, bt = b
return (aq * bq,
aq * br + ar * bt,
as_ * bq + at * bs,
as_ * br + at * bt)
def extract(z, j):
q, r, s, t = z
return (q * j + r) // (s * j + t)
def gen_pi_digits():
z = (1, 0, 0, 1)
x = gen_x()
while 1:
y = extract(z, 3)
while y != extract(z, 4):
z = compose(z, next(x))
y = extract(z, 3)
z = compose((10, -10 * y, 0, 1), z)
yield y
def calc_ndigits(n=DEFAULT_DIGITS):
return list(islice(gen_pi_digits(), n))
calc_ndigits()
"""
def bench_threading(n):
# Code to launch specific model
threads = []
for _ in range(n):
t = Thread(target=calc_ndigits)
t.start()
threads.append(t)
for thread in threads:
thread.join()
def bench_subinterpreters(n, site=True):
# Code to launch specific model
def _spawn_sub():
sid = interpreters.create()
interpreters.run_string(sid, test)
interpreters.destroy(sid)
threads = []
for _ in range(n):
t = Thread(target=_spawn_sub)
t.start()
threads.append(t)
for thread in threads:
thread.join()
def bench_multiprocessing(n):
# Code to launch specific model
processes = []
for _ in range(n):
t = Process(target=calc_ndigits)
t.start()
processes.append(t)
for process in processes:
process.join()
if __name__ == "__main__":
runner = pyperf.Runner()
runner.metadata['description'] = "Benchmark execution models"
n = 10
runner.bench_func('threading', bench_threading, n)
runner.bench_func('subinterpreters', bench_subinterpreters, n)
runner.bench_func('multiprocessing', bench_multiprocessing, n)
Results when running a CPython without --disable-gil:
Running the benchmark with PYTHON_GIL=0:
Running the benchmark with a no-gil build, but PYTHON_GIL=1 uses all 4 CPU cores and gives the fastest result (faster than PYTHON_GIL=0) which is wrong--
Okay this should be fixed, since I am busy before PyCon US, if you want to submit the patch before, you can send it :) Or I will take a look at it during the PyCon US.
I'd suggest looking at:
https://github.com/psf/pyperf/blob/555a16885096072eecd07b2a765d6f437add924c/pyperf/_utils.py#L270
IIRC, pyperf prevents workers from inheriting environment variables by default, which I've always found more confusing than helpful.
-Xgil is probably also not propagated.
