借鑒秦總的文章
https://mp.weixin.qq.com/s/WtozWOzNWlIJnWmcVUpYJQ

Nautius的分布調(diào)優(yōu)，方法略有不同。

1:開始調(diào)優(yōu)前，集群是10個1T ssd，兩個hdd。pg分布情況是。

image.png

除了HDD的，ssd的最大最小差34個pg，還是極其不均衡的。

2:開始調(diào)優(yōu)

nautilus改變了balancer module 此module是默認(rèn)永遠開啟的，所以我們不需要設(shè)置開啟此module

  $ module 'balancer' is already enabled (always-on)

   啟動balancer服務(wù)
  $ ceph balancer on
   設(shè)置crush-compat
  $ ceph balancer mode crush-compat
   (Luminous的max_misplaced在Nautilus已經(jīng)不存在了)
   看下狀態(tài)
  $ ceph balancer status
    {
       "active": true, 
       "plans": [], 
       "mode": "crush-compat"
     }
  對集群所有pool進行權(quán)重調(diào)整計算
  $ ceph balancer eval
  current cluster score 0.081403 (lower is better)
  生產(chǎn)一個調(diào)優(yōu)配置記住n版本必須制定pool
  $ ceph balancer optimize plan1 ssd
 開始執(zhí)行調(diào)優(yōu)計劃
  $ ceph balancer eval plan1
   plan plan1 final score 0.010106 (lower is better)
 查看調(diào)優(yōu)結(jié)果
  $ ceph balancer show plan1

         # starting osdmap epoch 19047                          
         # starting crush version 113                           
         # mode crush-compat                                    
         ceph osd crush weight-set reweight-compat 0 0.959158   
         ceph osd crush weight-set reweight-compat 1 0.886730   
         ceph osd crush weight-set reweight-compat 2 0.924439   
         ceph osd crush weight-set reweight-compat 3 0.830849   
         ceph osd crush weight-set reweight-compat 4 0.981423   
         ceph osd crush weight-set reweight-compat 5 0.454987   
         ceph osd crush weight-set reweight-compat 6 0.454987   
         ceph osd crush weight-set reweight-compat 7 0.968349   
         ceph osd crush weight-set reweight-compat 8 0.910090   
         ceph osd crush weight-set reweight-compat 9 0.963822   
         ceph osd crush weight-set reweight-compat 10 0.943591  
         ceph osd crush weight-set reweight-compat 11 0.941545  

 根據(jù)計劃進行最終調(diào)優(yōu)
 $ ceph balancer execute plan1
 調(diào)優(yōu)結(jié)果不滿意，回滾繼續(xù)重新來一遍(我執(zhí)行后發(fā)現(xiàn)效果不好，重新來一遍)
 $ ceph balancer reset

經(jīng)過多次嘗試得到完美結(jié)果

image.png

    關(guān)閉balancer模塊
    $ ceph balancer off

3:后記

pg分布這個重要性就不說了，建議在集群上線前必須要進行這個操作。對集群的整體性能提升絕對是有效果的。