[Linux-cluster] optimising DLM speed?
ajb2 at mssl.ucl.ac.uk
Thu Feb 24 18:37:24 UTC 2011
Steven Whitehouse wrote:
> As soon as you mix creation/deletion on one node with accesses (of
> whatever kind) from other nodes, you run this risk.
_ALL_ the GFS2 filesystems (bar one 5Gb one for common config files,
etc) are mounted one-node-only.
_ALL_ the GFS2 filesystems (with the same exception) are NFS exported.
NONE of the NFS exported filesystems have local processes accessing them
except for backups(*) - because there's distinct and non-theoretrical
risk of file coruption if anything other than NFSd touches a
NFS-exported filesystem (We've experienced it and I've reproduced the
corruption on non-cluster systems).
Even Samba is a re-export from a NFS client and I've been toying with
the idea of moving backups to a NFS client despite the network penalties.
(*) Backups run on the node where the filesystem is NFS exported.
> Obviously you wouldn't be using a cluster filesystem if you didn't intend to have this
> kind of access from time to time, but anything that can be done at the
> application level to help improve locality will pay big dividends
> compared with any tuning that can be done at the fs/dlm level.
We originally installed this to run as pNFS/SAMBA/iscsi fileservers but
after encountering the NFS corruption issues and finding out just how
much slower it gets if other nodes mount/access the filesystems we just
use GFS to ensure corruption-free failover.
Let's just say that what was promised was not what was delivered...
More information about the Linux-cluster