首页 > 解决方案 > 大型数组乘法后,进程关联性转到 {0}

问题描述

系统信息:

软件信息:

由于尝试在 1M 记录数据集上并行拟合sklearn.ensemble.RandomForestClassifiern_jobs=-1 ( ),我对此进行了追踪。安装后,我注意到joblib后端一次恢复到SequentialBackend正在运行的 1 个作业。

我确定在调用fit进程亲和性(如 所示os.sched_getaffinity)时已将其设置为单核,而不是正常的 24。事实证明,在数据转换中有一个步骤是 anumpy.array乘以 a float。在这一点上,OpenMP 抛出了一系列关于 KMP_AFFINITY 的消息,这些消息似乎都没有表明进程已设置为单核,但我不是 OpenMP 日志记录方面的专家。下面是一个在我的系统上重现此错误的简单示例。

import numpy as np
from os import sched_getaffinity as aff

aff(0)
{0,
 1,
 2,
 3,
 4,
 5,
 6,
 7,
 8,
 9,
 10,
 11,
 12,
 13,
 14,
 15,
 16,
 17,
 18,
 19,
 20,
 21,
 22,
 23}
arr = np.random.rand(1000000)
arr *= .5
OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.
OMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info
OMP: Info #154: KMP_AFFINITY: Initial OS proc set respected: 0-23
OMP: Info #156: KMP_AFFINITY: 24 available OS procs
OMP: Info #157: KMP_AFFINITY: Uniform topology
OMP: Info #179: KMP_AFFINITY: 24 packages x 1 cores/pkg x 1 threads/core (24 total cores)
OMP: Info #214: KMP_AFFINITY: OS proc to physical thread map:
OMP: Info #171: KMP_AFFINITY: OS proc 0 maps to package 0
OMP: Info #171: KMP_AFFINITY: OS proc 1 maps to package 2
OMP: Info #171: KMP_AFFINITY: OS proc 2 maps to package 4
OMP: Info #171: KMP_AFFINITY: OS proc 3 maps to package 6
OMP: Info #171: KMP_AFFINITY: OS proc 4 maps to package 8
OMP: Info #171: KMP_AFFINITY: OS proc 5 maps to package 10
OMP: Info #171: KMP_AFFINITY: OS proc 6 maps to package 12
OMP: Info #171: KMP_AFFINITY: OS proc 7 maps to package 14
OMP: Info #171: KMP_AFFINITY: OS proc 8 maps to package 16
OMP: Info #171: KMP_AFFINITY: OS proc 9 maps to package 18
OMP: Info #171: KMP_AFFINITY: OS proc 10 maps to package 20
OMP: Info #171: KMP_AFFINITY: OS proc 11 maps to package 22
OMP: Info #171: KMP_AFFINITY: OS proc 12 maps to package 24
OMP: Info #171: KMP_AFFINITY: OS proc 13 maps to package 26
OMP: Info #171: KMP_AFFINITY: OS proc 14 maps to package 28
OMP: Info #171: KMP_AFFINITY: OS proc 15 maps to package 30
OMP: Info #171: KMP_AFFINITY: OS proc 16 maps to package 32
OMP: Info #171: KMP_AFFINITY: OS proc 17 maps to package 34
OMP: Info #171: KMP_AFFINITY: OS proc 18 maps to package 36
OMP: Info #171: KMP_AFFINITY: OS proc 19 maps to package 38
OMP: Info #171: KMP_AFFINITY: OS proc 20 maps to package 40
OMP: Info #171: KMP_AFFINITY: OS proc 21 maps to package 42
OMP: Info #171: KMP_AFFINITY: OS proc 22 maps to package 44
OMP: Info #171: KMP_AFFINITY: OS proc 23 maps to package 46
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 8822 thread 0 bound to OS proc set 0
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9374 thread 1 bound to OS proc set 1
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9375 thread 2 bound to OS proc set 2
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9376 thread 3 bound to OS proc set 3
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9377 thread 4 bound to OS proc set 4
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9378 thread 5 bound to OS proc set 5
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9380 thread 7 bound to OS proc set 7
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9379 thread 6 bound to OS proc set 6
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9381 thread 8 bound to OS proc set 8
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9383 thread 10 bound to OS proc set 10
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9382 thread 9 bound to OS proc set 9
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9384 thread 11 bound to OS proc set 11
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9385 thread 12 bound to OS proc set 12
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9386 thread 13 bound to OS proc set 13
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9387 thread 14 bound to OS proc set 14
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9388 thread 15 bound to OS proc set 15
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9389 thread 16 bound to OS proc set 16
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9390 thread 17 bound to OS proc set 17
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9391 thread 18 bound to OS proc set 18
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9392 thread 19 bound to OS proc set 19
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9394 thread 21 bound to OS proc set 21
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9393 thread 20 bound to OS proc set 20
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9395 thread 22 bound to OS proc set 22
OMP: Info #250: KMP_AFFINITY: pid 8822 tid 9396 thread 23 bound to OS proc set 23

aff(0)
{0}

对于如何解释这种行为,我有点茫然,甚至在os.sched_setaffinity之后重置进程亲和力也不会恢复使用所有内核的能力。

注意

当数组的大小是 1e5 而不是 1e6 时,不会发生这种情况。内存不是问题,因为该系统有 256GB。

标签: pythonnumpyopenmp

解决方案


我在 OpenMP bugzilla 页面上的 Multiplication of array with float 下打开了一张票会导致 OpenMP 消息的级联并将进程关联设置为 {0}

我从一位英特尔工程师那里得到的建议是使用norespect KMP_AFFINITY修饰符。可在英特尔 C++ 编译器开发人员指南 - 线程关联接口中找到有关此修饰符和其他修饰符的文档。这对我有用。

export KMP_AFFINITY=norespect

ipython
In [1]: import numpy as np
In [2]: from os import sched_getaffinity as aff
In [3]: aff(0)
Out[3]:
{0,
 1,
 2,
 3,
 4,
 5,
 6,
 7,
 8,
 9,
 10,
 11,
 12,
 13,
 14,
 15,
 16,
 17,
 18,
 19,
 20,
 21,
 22,
 23}
In [4]: arr = np.random.rand(1000000)
In [5]: arr *= .5
OMP: Info #212: KMP_AFFINITY: decoding x2APIC ids.
OMP: Info #210: KMP_AFFINITY: Affinity capable, using global cpuid leaf 11 info
OMP: Info #155: KMP_AFFINITY: Initial OS proc set not respected: 0-23
OMP: Info #156: KMP_AFFINITY: 24 available OS procs
OMP: Info #157: KMP_AFFINITY: Uniform topology
OMP: Info #179: KMP_AFFINITY: 24 packages x 1 cores/pkg x 1 threads/core (24 total cores)
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4274 thread 0 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4697 thread 3 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4698 thread 4 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4699 thread 5 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4700 thread 6 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4701 thread 7 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4702 thread 8 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4703 thread 9 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4704 thread 10 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4705 thread 11 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4706 thread 12 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4708 thread 14 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4709 thread 15 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4710 thread 16 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4711 thread 17 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4712 thread 18 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4714 thread 20 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4715 thread 21 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4716 thread 22 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4717 thread 23 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4713 thread 19 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4707 thread 13 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4695 thread 1 bound to OS proc set 0-23
OMP: Info #250: KMP_AFFINITY: pid 4274 tid 4696 thread 2 bound to OS proc set 0-23
In [6]: aff(0)
Out[6]:
{0,
 1,
 2,
 3,
 4,
 5,
 6,
 7,
 8,
 9,
 10,
 11,
 12,
 13,
 14,
 15,
 16,
 17,
 18,
 19,
 20,
 21,
 22,
 23}

推荐阅读