[Linux-cluster] RE: < fecing with out any hardware? >
lhh at redhat.com
Mon May 8 13:49:21 UTC 2006
Sorry for the late response.
On Mon, 2006-05-01 at 21:45 +0200, Pool Lee, Mr <14117614 at sun.ac.za>
> What about software fencing? Is it really nesasary to be hardware!
Fencing basically is using a device which is not directly controlled by
cluster nodes to ensure a given node is cut off from performing I/O,
thereby corrupting shared data.
> Is there a difference between lutre/cfs, the product that sun uses, and gfs?
I have not read much about Sun's product(s), but GFS is significantly
different architecturally from Lustre.
GFS has no metadata or data servers per se (though, when using gulm, you
have a 'lock' server); all nodes are accessing the same block devices
> I'm planning to do mostly numerical work with the cluster and thus I would like all the machines to be able to
> retrieve data, as if it was local on the machine. NFS is very limited in this regard because we intend on using vast arrays of matrices, that can be up to 1-2 Gig.
You can use GFS and export the same NFS volume from multiple servers if
you need to, which helps eliminate the single-NFS-server bottleneck.
(In this case, you only need to set up fencing for the GFS cluster.)
Or, you can connect all the nodes in your cluster to the same block
devices and use GFS directly.
> I was hoping to implement GFS since all the machines are already setup, without the hardware fencing though.
Well, you /can/ do this, but if a node hangs and comes back to life,
plan on rebooting the entire cluster, recreating the file system from
scratch, and restarting your calculations.
More information about the Linux-cluster