Optimizing parsing of massive python dictionary, m

问题:

Let's take a small example python dictionary, where the values are lists of integers.

example_dict1 = {'key1':[367, 30, 847, 482, 887, 654, 347, 504, 413, 821],
    'key2':[754, 915, 622, 149, 279, 192, 312, 203, 742, 846], 
    'key3':[586, 521, 470, 476, 693, 426, 746, 733, 528, 565]}

Let's say I need to parse the values of the lists, which I've implemented into the following function:

def manipulate_values(input_list):
    return_values = []
    for i in input_list:
        new_value = i ** 2 - 13
        return_values.append(new_value)
    return return_values

Now, I can easily parse the values of this dictionary as follows:

for key, value in example_dict1.items():
    example_dict1[key] = manipulate_values(value)

resulting in the following:

example_dict1 = {'key1': [134676, 887, 717396, 232311, 786756, 427703, 120396, 254003, 170556, 674028], 
     'key2': [568503, 837212, 386871, 22188, 77828, 36851, 97331, 41196, 550551, 715703], 
     'key3': [343383, 271428, 220887, 226563, 480236, 181463, 556503, 537276, 278771, 319212]}

That works very well for small dictionaries.

My problem is, I have a massive dictionary with millions of keys and long lists. If I were to apply the above approach, the algorithm would be prohibitively slow.

How could I optimize the above?

(1) Multithreading---are there more efficient options available for multithreading this for statement in the dictionary besides the traditional threading module?

(2) Would a better data structure be appropriate?

I'm asking this question as, I'm quite stuck how to best proceed in this case. I don't see a better data structure than a dictionary, but the for loops across the dictionary (and then across the value lists) is quite slow. There may be something here which has been designed to be faster.

EDIT: As you can imagine, this is somewhat of a toy example---the function in question is a bit more complicated than x**2-13.

I'm more interested in how to possibly worth with a dictionary with millions of keys, with long lists of values.

回答1:

If you can store everything inside a numpy array processing will be faster. I increased the size of each list by a factor of 0.5 millions to test scalability, and these are my results:

from timeit import timeit
import numpy as np

n = 500000
example_dict1 = {'key1':[367, 30, 847, 482, 887, 654, 347, 504, 413, 821]*n,
    'key2':[754, 915, 622, 149, 279, 192, 312, 203, 742, 846]*n, 
    'key3':[586, 521, 470, 476, 693, 426, 746, 733, 528, 565]*n}

def manipulate_values(input_list):
    return_values = []
    for i in input_list:
        new_value = i ** 2 - 13
        return_values.append(new_value)
    return return_values

With your method:

for_with_dictionary = timeit("""
for key, value in example_dict1.items():
    example_dict1[key] = manipulate_values(value)
""", "from __main__ import example_dict1,manipulate_values ",number=5)

print(for_with_dictionary)

>>> 33.2095841

With numpy:

numpy_broadcasting = timeit("""
array = np.array(list(example_dict1.values()))
array = array ** 2 - 13
""", "from __main__ import example_dict1, np",number=5)
print(numpy_broadcasting)

>>> 5.039885

There is a significant upgrade in speed, at least 6 times.

回答2:

If you have enough RAM:

example_dict2 = dict(zip(example_dict1.keys(), np.array(list(example_dict1.values()))**2 -13))
>>> example_dict2
{'key1': array([134676,    887, 717396, 232311, 786756, 427703, 120396, 254003,
       170556, 674028]), 'key2': array([568503, 837212, 386871,  22188,  77828,  36851,  97331,  41196,
       550551, 715703]), 'key3': array([343383, 271428, 220887, 226563, 480236, 181463, 556503, 537276,
       278771, 319212])}