如果进程数超过内核数的一半,为什么性能会下降?

大家好,我有下面的Python程序,用于测试多处理应用程序的性能。

#
# Date : 09/May/2018
# Platform : Linux
#

import os
import sys
import ctypes
import signal
import multiprocessing as mp

ncpu = 4
counter = 0
child_index = 0
process_list = []
shared_array = None

def HandleSignal(signum, frame) :

   total = 0
   print("Parent timeout hence terminate child")
   [hProc.terminate() for hProc in process_list]
   [hProc.join() for hProc in process_list]
   for each_count in shared_array :
      total += each_count
   print("{:,}".format(total))

def ChildHandleSignal(signum, frame) :

   # print("{} - {} : {:,}".format(child_index, os.getpid(), counter))
   shared_array[child_index] = counter
   sys.exit(0)

def entry_point(index, sarr) :

   global counter
   global child_index
   global shared_array

   child_index = index
   shared_array = sarr
   signal.signal(signal.SIGTERM, ChildHandleSignal)
   while True : counter += 1

   return

ncpu = int(sys.argv[1])
maxcpu = os.cpu_count()

if ncpu > maxcpu :

   print("Number of CPU greater than maximum CPU")
   print("Setting number of CPU to maximum")
   ncpu = maxcpu

shared_array = mp.Array(ctypes.c_int64, range(ncpu))
signal.signal(signal.SIGALRM, HandleSignal)
signal.alarm(5)

for I in range(ncpu) :

   p1 = mp.Process(target=entry_point, args=(I, shared_array, ))
   process_list.append(p1)
   p1.start()

   # I tried both with and with-out the below
   # statement. The outputs are much similar
   os.sched_setaffinity(p1.pid, {I})

我在两台不同的机器上运行了这个程序

1. 使用Cent OS 7的谷歌云虚拟机。x在8 VCPU Intel处理器上运行
2. Cent OS 7。具有48核Intel处理器的X Linux计算机

输出与所用芯数的关系图如下所示。从中我观察到输出增加,直到进程数达到核数/2,然后下降。有人能解释一下这种行为吗?