numa node cpu分配不均衡

由于nova flavor 可以指定虚拟机cpu在numa node 使用跨node指定的方式，以及单个node指定的方式，后者性能会更好，但是这两种逻辑混用会出现 numa node不均衡的结果，导致双node flavor无法创建出虚拟机，即使总体来看，还有虚拟机剩余

基于 nova compute_nodes表统计 物理cpu的分配数目，手动对不均衡的虚拟机进行热迁移或者调整规格

# tools/get_numa_nodes.py
#!/usr/bin/env python

import mysql.connector
from pprint import pprint as pp
import json

def main():
    mydb = mysql.connector.connect(
      host="10.120.12.21",
      user="root",
      password="AVL13CYlSZvAYDTzYpJtu8Ykt2mWKjgf3CHcTHFW"
    )
    mycursor = mydb.cursor()
    
    mycursor.execute("SELECT host,numa_topology FROM nova.compute_nodes")
    
    myresult = mycursor.fetchall()
    numa_hosts = {}
    for x in myresult:
      host_name = x[0]
      numa_host_info = json.loads(x[1])
      #pp(numa_host_info)
      cells = numa_host_info['nova_object.data']['cells']
      #print('%s nodes: %s' % (host_name, len(cells)))
      #pp(cells)
      numa_hosts[host_name] = {}
      node0 = cells[0]['nova_object.data']
      node1 = cells[1]['nova_object.data']
      #numa_hosts[host_name]['node0'] =  node0
      #numa_hosts[host_name]['node1'] =  node1
      
      #numa_hosts[host_name]['node0'] = {}
      #numa_hosts[host_name]['node1'] = {}
      
      #numa_hosts[host_name]['node0']['pcpu_total'] = len(node0['pcpuset'])
      #numa_hosts[host_name]['node0']['pcpu_pined'] = len(node0['pinned_cpus'])
      
      numa_hosts[host_name]['node0'] = "%s/%s" % (len(node0['pinned_cpus']), len(node0['pcpuset']))
      numa_hosts[host_name]['node1'] = "%s/%s" % (len(node1['pinned_cpus']), len(node1['pcpuset']))


      #numa_hosts[host_name]['node1']['pcpu_total'] = len(node1['pcpuset'])
      #numa_hosts[host_name]['node1']['pcpu_pined'] = len(node1['pinned_cpus'])

    pp(numa_hosts)    
    
if __name__ == '__main__':
    main()

基于virsh 定位到不均衡的机器

(nova-libvirt)[root@compute002 /]# virsh list --all --uuid
88feed61-d3aa-4dd6-b752-1a2760f0b5c9
b88c197f-24bb-4632-96c3-7183f274674f
260aa0aa-c97d-42dd-890e-4abd2cd644a1
dde5564a-59e1-45e1-85a5-acbee4b29b92
3c07a3ca-009e-4688-9f24-08e57b6aa115
7cdab948-065a-47b6-9f5d-4edf18131e20
2f7e51ab-2983-47cd-be45-1380ec121abc
ecdea0d0-9aa8-4f98-bec1-eadf20bbfa3b
8c7d8023-25d1-42c4-9062-d802bd4540cb


(nova-libvirt)[root@compute002 /]# virsh numatune 88feed61-d3aa-4dd6-b752-1a2760f0b5c9
numa_mode      : strict
numa_nodeset   : 

共享型虚拟机
(nova-libvirt)[root@compute002 /]# virsh numatune b88c197f-24bb-4632-96c3-7183f274674f
numa_mode      : strict
numa_nodeset   : 0

cpu 只在numa 0 上面


(nova-libvirt)[root@compute002 /]# virsh numatune 260aa0aa-c97d-42dd-890e-4abd2cd644a1
numa_mode      : strict
numa_nodeset   : 0-1

cpu 平均占用在两个numa node上

然后逐个处理即可