Memory Footprint

Decimals 比 floats 要佔用更多記憶體：

import sys
from decimal import Decimal

a = 3.1415
b = Decimal('3.1415')

sys.getsizeof(a)

24

儲存 float 3.1415 需 24 bytes

sys.getsizeof(b)

104

然而 Decimal 3.1415 需要佔用 104 bytes。

Computational Performance

Decimal 的數學運算也比 float 慢很多 (聽說在 GPU 上比 CPU 更明顯)

import time
from decimal import Decimal

def run_float(n=1):
    for i in range(n):
        a = 3.1415
        
def run_decimal(n=1):
    for i in range(n):
        a = Decimal('3.1415')

用上面兩個函式來計算兩者的運算速度：

n = 10000000

start = time.perf_counter()
run_float(n)
end = time.perf_counter()
print('float: ', end-start)

start = time.perf_counter()
run_decimal(n)
end = time.perf_counter()
print('decimal: ', end-start)

float:  0.12886674999026582
decimal:  0.884507458016742

可以看到生成一個 Decimal 物件遠比 float 物件花時間。

那加減乘除等等運算呢？來測試一下：

def run_float(n=1):
    a = 3.1415
    for i in range(n):
        a + a
        
def run_decimal(n=1):
    a = Decimal('3.1415')
    for i in range(n):
        a + a
        
start = time.perf_counter()
run_float(n)
end = time.perf_counter()
print('float: ', end-start)

start = time.perf_counter()
run_decimal(n)
end = time.perf_counter()
print('decimal: ', end-start)

float:  0.18404087499948218
decimal:  0.3708067910047248

速度差了兩倍左右。

最後來測試一下開根號:

(把 n 調低一點，因為⋯⋯)

n = 5000000

import math

def run_float(n=1):
    a = 3.1415
    for i in range(n):
        math.sqrt(a)
        
def run_decimal(n=1):
    a = Decimal('3.1415')
    for i in range(n):
        a.sqrt()
        
start = time.perf_counter()
run_float(n)
end = time.perf_counter()
print('float: ', end-start)

start = time.perf_counter()
run_decimal(n)
end = time.perf_counter()
print('decimal: ', end-start)

float:  0.20593712502159178
decimal:  6.594072999956552

Decimal 很慢，所以使用前也要三思，通常是為了精準度才會考慮用它。

好啦，我們明天見～

參考：Python 3: Deep Dive (Part 1 - Functional)