test_multiprocessing.py

from __future__ import division, print_function

from concurrent import futures
import logging
import math
import multiprocessing
import os
import pickle
import platform
import pprint
import random
import time
import timeit
import unittest

import cloudpickle

import pysparkling


class Processor(object):
    """This modifies lines but also keeps track whether it was executed."""
    def __init__(self):
        self.executed = False

    def indent_line(self, line):
        self.executed = True
        return '--- {}'.format(line)


class LazyTestInjection(object):
    def lazy_execution_test(self):
        r = self.sc.textFile(__file__)  # pylint: disable=no-member

        processor = Processor()

        r = r.map(processor.indent_line)
        self.assertFalse(processor.executed)  # pylint: disable=no-member
        r = r.map(processor.indent_line).cache()
        self.assertFalse(processor.executed)  # pylint: disable=no-member
        r = r.map(processor.indent_line)
        r.collect()
        self.assertTrue(processor.executed)  # pylint: disable=no-member


class Multiprocessing(unittest.TestCase):
    def setUp(self):
        self.pool = multiprocessing.Pool(4)
        self.sc = pysparkling.Context(pool=self.pool,
                                      serializer=cloudpickle.dumps,
                                      deserializer=pickle.loads)

    def test_basic(self):
        my_rdd = self.sc.parallelize([1, 3, 4])
        r = my_rdd.map(lambda x: x ** 2).collect()
        self.assertIn(16, r)

    def test_first(self):
        my_rdd = self.sc.parallelize([1, 2, 2, 4, 1, 3, 5, 9], 3)
        self.assertEqual(my_rdd.first(), 1)

    def tearDown(self):
        self.pool.close()


def square_op(x):
    return x ** 2


class MultiprocessingWithoutCloudpickle(unittest.TestCase):
    def setUp(self):
        self.pool = multiprocessing.Pool(4)
        self.sc = pysparkling.Context(pool=self.pool)

    def test_basic(self):
        my_rdd = self.sc.parallelize([1, 3, 4])
        r = my_rdd.map(square_op).collect()
        self.assertIn(16, r)

    def tearDown(self):
        self.pool.close()


class NotParallel(unittest.TestCase, LazyTestInjection):
    """Test cases in the spirit of the parallel test cases for reference."""

    def setUp(self):
        self.sc = pysparkling.Context()


class ThreadPool(unittest.TestCase, LazyTestInjection):
    def setUp(self):
        self.pool = futures.ThreadPoolExecutor(4)
        self.sc = pysparkling.Context(pool=self.pool)

    def tearDown(self):
        self.pool.shutdown()

    def test_basic(self):
        r = self.sc.parallelize([1, 3, 4]).map(math.sqrt).collect()
        self.assertIn(2, r)


class ProcessPool(unittest.TestCase):  # cannot work here: LazyTestInjection):
    def setUp(self):
        self.pool = futures.ProcessPoolExecutor(4)
        self.sc = pysparkling.Context(pool=self.pool,
                                      serializer=cloudpickle.dumps,
                                      deserializer=pickle.loads)

    def tearDown(self):
        self.pool.shutdown()

    def test_basic(self):
        r = self.sc.parallelize([1, 3, 4]).map(math.sqrt).collect()
        self.assertIn(2, r)

    def test_zipWithIndex(self):
        """Prevent regression in zipWithIndex().

        Test the case of parallelizing data directly form toLocalIterator()
        in the multiprocessing case.
        """
        r = (self.sc
             .parallelize([1, 3, 4, 9, 15, 25, 50, 75, 100], 3)
             .zipWithIndex()
             .collect())
        self.assertIn((4, 2), r)

    def test_cache(self):
        r = self.sc.parallelize(range(3), 3)

        def sleep05(v):
            time.sleep(0.5)
            return v

        r = r.map(sleep05).cache()
        self.assertEqual(r.collect(), [0, 1, 2])

        start = time.time()
        r.collect()
        self.assertLess(time.time() - start, 0.5)


class ProcessPoolIdlePerformance(unittest.TestCase):
    """Idle performance tests.

    The "load" on these tests are sleeps.
    """

    def runtime(self, n=10, processes=1):
        start = time.time()
        with futures.ProcessPoolExecutor(processes) as pool:
            sc = pysparkling.Context(pool=pool,
                                     serializer=cloudpickle.dumps,
                                     deserializer=pickle.loads)
            rdd = sc.parallelize(range(n), 10)
            rdd.map(lambda _: time.sleep(0.1)).collect()
        return time.time() - start

    @unittest.skipIf(platform.python_implementation() == 'PyPy',
                     'test fails in PyPy')
    def test_basic(self):
        t1 = self.runtime(processes=1)
        t10 = self.runtime(processes=10)
        self.assertLess(t10, t1 / 2.0)


# pickle-able map function
def map1(ft):
    return [random.choice(ft[1].split()) for _ in range(1000)]


def map_pi(n):
    return sum((
        1 for x in (random.random() ** 2 + random.random() ** 2
                    for _ in range(n))
        if x < 1.0
    ))


@unittest.skipIf(os.getenv('PERFORMANCE') is None,
                 'PERFORMANCE env variable not set')
def test_performance():
    # not pickle-able map function
    # def map2(ft):
    #     return [random.choice(ft[1].split()) for _ in range(1000)]

    def create_context(n_processes=0):
        if not n_processes:
            return pysparkling.Context()

        pool = futures.ProcessPoolExecutor(n_processes)
        return pysparkling.Context(pool=pool,
                                   serializer=cloudpickle.dumps,
                                   # serializer=pickle.dumps,
                                   deserializer=pickle.loads)

    def test(n_processes):
        sc = create_context(n_processes)
        timed = timeit.Timer(
            lambda: sc.parallelize(
                [1000 for _ in range(100)],
                100,
            ).map(map_pi).collect()
        ).timeit(number=10)
        return (timed, sc._stats)

    print('starting processing')
    n_cpu = multiprocessing.cpu_count()
    test_results = {}
    for n in range(int(n_cpu * 1.5 + 1)):
        test_results[n] = test(n)
        print(n, test_results[n][0])
    print('results where running on one core with full serialization is 1.0:')
    pprint.pprint({
        n: 1.0 / (v[0] / test_results[1][0]) for n, v in test_results.items()
    })
    print('time spent where:')
    pprint.pprint({
        n: {k: '{:.1%}'.format(t / v[1]['map_exec']) for k, t in v[1].items()}
        for n, v in test_results.items()
    })

    return (n_cpu, test_results)


if __name__ == '__main__':
    logging.basicConfig(level=logging.INFO)
    # test_performance()
    t = ProcessPool()
    t.setUp()
    t.test_cache()
    t.tearDown()