Re: Crush maps : split the root in two parts on an OSD node with same disks ?

Hervé Ballans <herve.ballans@xxxxxxxxxxxxx> · Wed, 13 Jun 2018 11:08:48 +0200



    Thanks Janne for your reply.

      
      Here are the reasons which made me think to "physically" split the
      pools :

      
      1) A different usage of the pools : the first one will be used for
      user home directories, with an intensive read/write access. And
      the second one will be used for data storage/backup, with
      essentially a read access once data will be stored.

      
      2) Pools will be CephFS data pools. On each MON node I have 2 SSD
      NVMe dedicated for cephfs metadata pools. Here too, I thought
      about splitting the NVMe in the crush maps in order to have 2
      different metadata pools. To sum up, the first data pool is
      associated with a metadata pool composed of one NVMe per MON node,
      and the second data pool is associated with a metadata pool
      composed of the second NVMe of each MON.

      
      And a third reason :

      
      3) Since they are all identical disks, I understand that all disks
      are automatically put on the same root of the crush maps, right ?

      Actually, I don't know (and I thought it was not possible) how to
      create 2 pools from this default configuration ?..

      
      In fact, if the third point solve my issue, so it could indeed
      simplify my pool design !?..

      
      Regards,

      Hervé

      
      Le 12/06/2018 à 16:52, Janne Johansson a écrit :

    
          Den tis 12 juni 2018 kl 15:06 skrev Hervé
            Ballans <herve.ballans@xxxxxxxxxxxxx>:

          
          Hi all,

            
            I have a cluster with 6 OSD nodes, each has 20 disks, all of
            the 120 

            disks are strictly identical (model and size).

            (The cluster is also composed of 3 MON servers on 3 other
            machines)

            
            For design reason, I would like to separate my cluster
            storage into 2 

            pools of 60 disks.

            
            My idea is to modify the crushmap on each node in order to
            split the 

            root top hierarchy in two groups, ie 10 disks of each OSD
            node for the 

            first pool and the 10 others disks of each nodes for the
            second pool.

            
            I already did that on another cluster with 2 sets of disks
            of different 

            technology (HDD vs SSD) inspiring by : 

            https://www.sebastien-han.fr/blog/2014/08/25/ceph-mix-sata-and-ssd-within-the-same-box/

            
            But is it relevant to do that when we have a set of
            identical disks ?

          
          You could do it this way, but you could also just run two
            pools over the same 120 OSD disks.
           Perhaps if you stated the end goal you are trying to
            reach, it would be easier to figure out an
          answer to if it's relevant or not?
          

          The storage admin in my thinks you spread load and risk
            better if all 120 disks get used for
          both pools, but you might have a specific reason and if
            so, may we know it?
        
        
        -- 

        May the most significant bit
          of your life be positive.

        
_______________________________________________
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com