Predicting which OSD will fill up first

Predicting which OSD will fill up first

[Loic Dachary]

Hi Ceph,

Last week-end I wrote a script to show which hosts or OSD are the most overused. Then it occurred to me that it would be useful to a sysadmin wanting to keep an eye on the device usage. Some OSDs will fill up more quickly than others, there is no way around it (see http://tracker.ceph.com/issues/15653#detailed-explanation for the details) but that can easily be predicted and more space added to the cluster when necessary. For instance:

$ pip install crush
$ ceph osd crush dump > crushmap-ceph.json
$ crush ceph --convert crushmap-ceph.json > crushmap.json
$ crush analyze --type device --rule replicated --crushmap crushmap.json

                    ~id~  ~weight~  ~over/under used~
            ~name~
            osd.35    35  2.299988     10.400604
            osd.2      2  1.500000     10.126750
            osd.47    47  2.500000      5.543335
            osd.46    46  1.500000      2.956655
            osd.29    29  1.784988      2.506855
            ...

shows that more disks are needed when osd.35 reaches the threshold (80% full for instance). There is a blog post with more details at http://dachary.org/?p=3980

Cheers 
-- 
Loïc Dachary, Artisan Logiciel Libre