Python进程池multiprocessing.Pool

环境：

鲲鹏920:192核心
内存：756G
python：3.9

python单进程的耗时

在做单纯的cpu计算的场景，使用单进程核多进程的耗时做如下测试：
单进程情况下cpu的占用了如下，占用一半的核心数：

每一步和总耗时如下：

多进程

cpu占用如下，每个进程基本占用48个左右核心数；

多进程的耗时如下：
每一个进程的耗时为63s左右，总的耗时比单进程还多，如果绑定48核心到每个进程，耗时更高。这是为何？

是否可以得出结论，在cpu计算密集的场景，单进程（每个任务都是独立的、排除IO、竞争关系）的效率会比多进程会高呢?
注：同样的代码在x86服务器上测试过，结论依旧是单进程耗时比多进程会少，这是为什么？

样例代码

from sklearn.datasets import load_wine
from sklearn.preprocessing import MinMaxScaler, Normalizer, StandardScaler, RobustScaler
from sklearn.neural_network import MLPClassifier
from sklearn.model_selection import train_test_split
import time
from multiprocessing import Process, Pool, current_process
import multiprocessing
import numpy as np
import os
import psutil

import os

core_count = os.cpu_count()
print(f"The CPU has {
      
      core_count} cores.")

cpu_cores = [index for index in range(0, core_count)]

def task1(data):
    start = time.time()
    X = np.random.rand(178, 13)
    y = np.random.randint(low=0, high=3, size=(178))
    X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=60)
    mm_scaler = MinMaxScaler()
    X_train = mm_scaler.fit_transform(X_train)
    X_test = mm_scaler.fit_transform(X_test)

    mlp = MLPClassifier(solver='lbfgs', hidden_layer_sizes=[500, 500], max_iter=300, random_state=60)
    mlp.fit(X_train, y_train)

    # print("***" * 10, "current data value:{}".format(data))
    # print("******************************************current processid:{} end id is {}".format(multiprocessing.current_process().name, data))
    print("this step spend time is {} seconds".format(time.time() - start))
    # time.sleep(5)


def task(data):
    process = current_process()
    print(process)
    pid = os.getpid()
    index = process._identity[0]

    cores = cpu_cores[(index-1) * 48 : index * 48]
    # print("process:{}, pid:{}, index:{}, core:{}".format(process, pid, index, cores))
    p = psutil.Process(pid)  # 通过进程 ID 获取进程对象
    # p.cpu_affinity(cores)  # 绑定核心


    start = time.time()
    X = np.random.rand(178, 13)
    y = np.random.randint(low=0, high=3, size=(178))
    X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=60)
    mm_scaler = MinMaxScaler()
    X_train = mm_scaler.fit_transform(X_train)
    X_test = mm_scaler.fit_transform(X_test)

    mlp = MLPClassifier(solver='lbfgs', hidden_layer_sizes=[500, 500], max_iter=300, random_state=60)
    mlp.fit(X_train, y_train)

    print("this step spend time is {} seconds".format(time.time() - start))

def main():
    data = [i for i in range(4)]

    start = time.time()
    for item in data:
       task1(item)
    print("single spend time is ", time.time() - start, " seconds")
    start = time.time()
    with Pool(4) as pool:
        pool.map_async(task, data)
        pool.close()
        pool.join()
    print("spend time is ", time.time() - start, " seconds")

if __name__ == '__main__':
    main()