[Beowulf] scaling / file serving

[John Hearns]

On Thu, 2004-06-10 at 22:55, Joe Landman wrote:

> 
> Building a disk system is easy, building one that scales well is
hard.  
> 
> The other issue is local disk.  There are some folks absolutely
> horrified at the prospect of a cluster node having a local disk. 
Makes
> management harder.  
I fully realise there are arguments both ways.
We normally have a local disk on cluster nodes. It is formatted at
install time, and either synced or formatted at reboot time.
We can choose a partition table on a per-node basis.
The clusters we put in for the NGS in the UK have two disks per node,
one for data as you describe. We don't software RAID them though!

I heard at CERN that if you are buying ATX cased systems it was more
expensive to have them without disks in large quantities.
(Let's not ignite the rack/ATX debate please)