[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Linux-cluster] Simplest 4 node GFS 6.1 cluster

On Thu, 2005-06-23 at 08:43 -0500, Troy Dawson wrote:

> This idea of fencing is what's throwing me off.  If I'm reading things 
> right, I can't do group GFS without them being in a cluster, and they 
> can't be in a cluster without doing fencing.  But the fencing seems to 
> just allow the various machines to take over for one another.

Actually - fencing prevents hung rogue nodes from being able to corrupt
the file system.
Believe it or not, power-cycle fencing actually can help more than you
think: A node without power can't flush buffers, so after the node is
fenced, you can have it turned back on.  If it was a software failure,
your cluster will resume all normal operations without any manual

> I also don't have access to the SAN switch, other than my machines plug 
> into it.  It's essentially a black box.  These machines also don't have 
> any way to remotely turn power on an off.


> Is GFS what I really want?  I've tried just standard ext3, but I was 
> getting a caching problem with my read only machines.  Do I just want to 
> try and fix my caching problem?

You'll probably need to do synchronous I/O on all nodes.  This will
likely be slow, but I think your limiting factor will be network
bandwidth, not disk I/O times.

Note that GFS was designed to prevent "hot spots": places on the disk
which are accessed over and over -- like an inode table on ext3, for
example.  Overuse of hot-spots can cause premature failure of drives.

Just things to consider.  You can probably do it without GFS, but I
wouldn't recommend it.  Remote power control does not have to be
expensive. E.g.:


That's cheaper than replacing ONE enterprise-grade SCSI or FC disk.

(Disclaimer: I have no affiliation whatsoever with the seller.)

-- Lon

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]