25 May
2021
25 May
'21
10:12 a.m.
Den tis 25 maj 2021 kl 02:51 skrev Tuffli, Chuck <chuck.tuffli(a)hpe.com>om>:
I found a warning in Red Hat documentation regarding
OSD and RBD:
Ceph Block Devices must be deployed on separate nodes from the Ceph Monitor and OSD
nodes. Running kernel clients and kernel server daemons on the same node can lead to
kernel deadlocks. [1]
It was hard to tell from the documentation if this is still true as some Rook
documentation [2] mentioned a solution. Is having an RBD on the same node as the OSD's
still a no-no or is it OK with the conditions listed in the Rook docs (i.e. patched kernel
+ no XFS)? TIA!
I think it warns against a machine hosting data on an OSD while at the
same time mounting an RBD which has parts served by the OSD it serves.
If you end up with a situation where both parts require lots of memory
(during OSD recovery for instance) and the RBD (or the filesystem on
top of it) also needs memory, stuff can wedge so both suffer. The same
would go for running a mon on the same machine as the OSD, if the OSD
memory requirements pushes the machine into swap for instance, then
OSD recovery would be slower but probably be "more or less ok", but if
the mon process gets swapped out at the same time, it will most
certainly not be a good member of the mon quorum and that will
potentially lead to even worse issues than a slow OSD recovering.
If all things are perfect, you have endless ram and cpu and iops and
so on then this might not apply to you, but if you are in the lower
50% of machine resources like half the ceph admins out here, then it
might just be a good idea to consider the above warning.
--
May the most significant bit of your life be positive.