6.2. Micro-benchmarking

We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil -- Donald Knuth

6.2.1. Evaluation

  • Fresh start of Python process

  • Clean memory before start

  • Same data

  • Same start conditions, CPU load, RAM usage, iostat

  • Do not measure how long Python wakes up

  • Check what you measure

6.2.2. timeit

6.2.2.1. Programmatic use

Listing 6.64. Timeit simple statement
from timeit import timeit


setup = """name = 'José Jiménez'"""
stmt = """result = f'My name... {name}'"""

duration = timeit(stmt, setup, number=10000)

print(duration)
# 0.0005737080000000061
Listing 6.65. Timeit multiple statements with setup code
from timeit import timeit


setup = """
firstname = 'José'
lastname = 'Jiménez'
"""

TEST = dict()
TEST[0] = 'name = f"{firstname} {lastname}"'
TEST[1] = 'name = "{0} {1}".format(firstname, lastname)'
TEST[2] = 'name = firstname + " " + lastname'
TEST[3] = 'name = " ".join([firstname, lastname])'


for stmt in TEST.values():
    duration = timeit(stmt, setup, number=10000)
    print(f'{duration:.5}\t{stmt}')

# 0.00071559    name = f"{firstname} {lastname}"
# 0.0026514     name = "{0} {1}".format(firstname, lastname)
# 0.001015      name = firstname + " " + lastname
# 0.0013494     name = " ".join([firstname, lastname])
Listing 6.66. Timeit with globals()
from timeit import timeit


def factorial(n: int) -> int:
    if n == 0:
        return 1
    else:
        return n * factorial(n-1)


duration = timeit(
    stmt='factorial(500); factorial(400); factorial(450)',
    globals=globals(),
    number=10000,
)

duration = round(duration, 6)

print(f'factorial time: {duration} seconds')
# factorial time: 2.845382 seconds

6.2.2.2. Console use

Listing 6.67. Timeit
python3 -m timeit -n100000 -r100 --setup='name="Jose Jimenez"' 'output = f"My name... {name}"'
# 100000 loops, best of 100: 55.9 nsec per loop

python3 -m timeit -n100000 -r100 --setup='name="Jose Jimenez"' 'output = "My name... {name}".format(name=name)'
# 100000 loops, best of 100: 327 nsec per loop

python3 -m timeit -n100000 -r100 --setup='name="Jose Jimenez"' 'output = "My name... %s" % name'
# 100000 loops, best of 100: 124 nsec per loop

-n N, --number=N
how many times to execute ‘statement’

-r N, --repeat=N
how many times to repeat the timer (default 5)

-s S, --setup=S
statement to be executed once initially (default pass)

-p, --process
measure process time, not wallclock time, using time.process_time() instead of time.perf_counter(), which is the default

-u, --unit=U
specify a time unit for timer output; can select nsec, usec, msec, or sec

-v, --verbose
print raw timing results; repeat for more digits precision

-h, --help
print a short usage message and exit

6.2.3. Case Studies - Unique Keys

  • Runtime: Jupyter %%timeit

Listing 6.68. Setup code used for all examples
DATA = [
    {'Sepal length': 5.1, 'Sepal width': 3.5, 'Species': 'setosa'},
    {'Petal length': 4.1, 'Petal width': 1.3, 'Species': 'versicolor'},
    {'Sepal length': 6.3, 'Petal width': 1.8, 'Species': 'virginica'},
    {'Sepal length': 5.0, 'Petal width': 0.2, 'Species': 'setosa'},
    {'Sepal width': 2.8, 'Petal length': 4.1, 'Species': 'versicolor'},
    {'Sepal width': 2.9, 'Petal width': 1.8, 'Species': 'virginica'},
]

6.2.3.1. Append if object not in the list

%%timeit -r 10 -n 1000000

fieldnames = list()

for row in DATA:
    for key in row.keys():
        if key not in fieldnames:
            fieldnames.append(key)

# 2.16 µs ± 26.5 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.2. Append to list and deduplicate at the end

%%timeit -r 10 -n 1000000

fieldnames = list()

for row in DATA:
    for key in row.keys():
        fieldnames.append(key)

set(fieldnames)

# 2.5 µs ± 32.9 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.3. Add to set

%%timeit -r 10 -n 1000000

fieldnames = set()

for row in DATA:
    for key in row.keys():
        fieldnames.add(key)

# 2.12 µs ± 32.4 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.4. Update set

%%timeit -r 10 -n 1000000

unique_keys = set()

for row in DATA:
    unique_keys.update(row.keys())

# 1.57 µs ± 26.7 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.5. Set Comprehension

%%timeit -r 10 -n 1000000

fieldnames = set(key
    for record in DATA
        for key in record.keys())

# 2.06 µs ± 79.7 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.6. Add to Set Comprehension

  • Code appends generator object not values, this is why it is so fast!

%%timeit -r 10 -n 1000000

fieldnames = set()

fieldnames.add(key
    for record in DATA
       for key in record.keys())

# 447 ns ± 9.52 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.7. Update Set Comprehension

%%timeit -r 10 -n 1000000

fieldnames = set()
fieldnames.update(tuple(x.keys()) for x in DATA)

# 2.06 µs ± 45.9 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.3.8. Others

%%timeit -r 10 -n 1000000

unique_keys = set()

for row in DATA:
    unique_keys.update(tuple(row))

# 2.09 µs ± 16.1 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

%%timeit -r 10 -n 1000000

unique_keys = set()

for row in DATA:
    unique_keys.update(list(row))

# 2.33 µs ± 30.2 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

%%timeit -r 10 -n 1000000

unique_keys = set()

for row in DATA:
    unique_keys.update(set(row))

# 1.71 µs ± 54 ns per loop (mean ± std. dev. of 10 runs, 1000000 loops each)

6.2.4. Case Study - Fibonacci

6.2.4.1. Without cache

%%timeit

def factorial_nocache(n: int) -> int:
    if n == 0:
        return 1
    else:
        return n * factorial_nocache(n - 1)


factorial_nocache(500)
factorial_nocache(400)
factorial_nocache(450)

# 283 µs ± 6.63 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

6.2.4.2. Cache contains

%%timeit

CACHE = {}

def factorial_cache(n: int) -> int:
    if n in CACHE:
        return CACHE[n]

    if n == 0:
        return 1
    else:
        result = CACHE[n] = n * factorial_cache(n-1)
        return result


factorial_cache(500)
factorial_cache(400)
factorial_cache(450)

# 153 µs ± 2.49 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

6.2.4.3. Cache get

%%timeit

CACHE = {}

def factorial_cache(n: int) -> int:
    result = CACHE.get(n)

    if result:
        return result

    if n == 0:
        return 1
    else:
        result = CACHE[n] = n * factorial_cache(n-1)
        return result


factorial_cache(500)
factorial_cache(400)
factorial_cache(450)

# 181 µs ± 10.3 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

6.2.4.4. Cache contains with exceptions

%%timeit

CACHE = {}

def factorial_cache(n: int) -> int:
    if n == 0:
        return 1

    try:
        return CACHE[n]
    except KeyError:
        CACHE[n] = result = n * factorial_cache(n-1)
        return result


factorial_cache(500)
factorial_cache(400)
factorial_cache(450)

# 618 µs ± 6.6 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

6.2.4.5. Adding cache layer

%%timeit

CACHE = {}

def factorial_1(n: int) -> int:

    def factorial(n: int) -> int:
        if n == 0:
            return 1
        return n * factorial(n-1)

    if not n in CACHE:
        CACHE[n] = factorial(n)

    return CACHE[n]


factorial_1(500)
factorial_1(400)
factorial_1(450)

# 283 µs ± 6.44 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

6.2.4.6. Get from cache

%%timeit

CACHE = {}

def factorial2(n: int) -> int:
    if n == 0:
        return 1

    if n in CACHE:
        return CACHE[n]

    result = CACHE[n] = n * factorial2(n-1)
    return result


factorial2(500)
factorial2(400)
factorial2(450)

# 153 µs ± 9.64 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)

_cache = {}


def factorial(n):
    if n == 0:
        return 1

    if (result := _cache.get(n)):
        return result

    result = n * factorial(n-1)
    _cache[n] = result
    return result