Numpy meshgrid函数间歇性性能骤降问题排查

反复调用Numpy的meshgrid函数时，发现该函数在特定样本量下会毫无征兆地出现性能骤降。测试代码如下：

from time import time
import numpy as np
np.random.seed(1729)

maxn  = 760
times = np.empty(maxn+1)
times[0] = time()
for n in range(1, maxn+1):
    x = np.random.uniform(size=(n,1))
    y = np.random.uniform(size=(n,1))
    x_grid, y_grid = np.meshgrid(x, y)
    times[n] = time()

last = 40
[f'{i+maxn-last+1}: {x:.3f}' for i, x in enumerate(np.diff(times)[-last:])]

运行后得到结果：

['721: 0.003',
 '722: 0.003',
 '723: 0.003',
 '724: 0.004',
 '725: 0.010',
 '726: 0.007',
 '727: 0.007',
 '728: 0.005',
 '729: 0.005',
 '730: 0.006',
 '731: 37.243', # Suddenly slow
 '732: 0.009',
 '733: 0.007',
 '734: 0.007',
 '735: 0.009',
 '736: 0.007',
 '737: 0.009',
 '738: 0.009',
 '739: 0.009',
 '740: 11.602', # Suddenly slow
 '741: 0.008',
 '742: 0.010',
 '743: 0.012',
 '744: 0.012',
 '745: 0.016',
 '746: 0.009',
 '747: 0.015',
 '748: 0.015',
 '749: 61.460', # Suddenly slow
 '750: 0.008',
 '751: 0.007',
 '752: 0.007',
 '753: 0.007',
 '754: 0.007',
 '755: 0.007',
 '756: 0.007',
 '757: 0.007',
 '758: 1.167', # Suddenly slow
 '759: 0.007',
 '760: 0.010']

可以看到，多数调用耗时极短，但部分特定调用（如示例中从731开始每隔9次调用）耗时骤增。不确定在其他机器上需要设置多大的maxn才能复现该现象，希望了解可能的原因。

环境信息：Arch Linux系统，Python 3.10.10，Numpy 1.24.2。

补充：perf stat测试结果 #

使用perf stat运行该代码片段后，性能骤降出现的位置有所不同：

721: 0.003
722: 0.003
723: 0.003
724: 0.004
725: 0.008
726: 0.009
727: 0.008
728: 0.009
729: 0.014
730: 0.015
731: 0.013
732: 0.009
733: 0.009
734: 16.755
735: 0.012
736: 0.011
737: 0.011
738: 0.011
739: 0.013
740: 0.010
741: 0.010
742: 0.013
743: 0.011
744: 22.704
745: 0.016
746: 0.012
747: 0.017
748: 0.016
749: 0.016
750: 0.012
751: 0.011
752: 0.016
753: 4.244
754: 0.018
755: 0.014
756: 0.018
757: 0.015
758: 0.017
759: 0.021
760: 0.018

 Performance counter stats for 'python test.py':

         45,125.79 msec task-clock:u                     #    0.998 CPUs utilized              
                 0      context-switches:u               #    0.000 /sec                       
                 0      cpu-migrations:u                 #    0.000 /sec                       
           568,187      page-faults:u                    #   12.591 K/sec                      
     1,150,132,500      cycles:u                         #    0.025 GHz                        
     1,124,532,475      instructions:u                   #    0.98  insn per cycle            
       188,859,225      branches:u                       #    4.185 M/sec                      
         3,959,138      branch-misses:u                  #    2.10% of all branches            

      45.194501699 seconds time elapsed

       0.393526000 seconds user
      44.537625000 seconds sys

补充2：meshgrid替代方案 #

已找到适用于该场景的meshgrid替代方案：

x_grid, y_grid = np.broadcast_arrays(x[:, np.newaxis], y[np.newaxis, :])
x_grid = x_grid.reshape((x.shape[0], y.shape[0])).T
y_grid = y_grid.reshape((x.shape[0], y.shape[0])).T

内容的提问来源于stack exchange，提问作者Mauricio