From: Jan-Frode Myklebust (janfrode@PARALLAB.UIB.NO)
Date: Wed Oct 02 2002 - 03:01:34 EDT
On Tue, Oct 01, 2002 at 12:15:05PM +0200, Green, Simon wrote:
> I'm pretty sure that GPFS is overkill for your situation.
>
> It was originally an SP2 thing, and required the HiP or SP Switch. More
> recent versions will also run over other high-performance networks and it's
> no longer restricted to SP2.
>
> The idea was that if you had a sufficiently fast network, you could split
> your I/O over multiple nodes, taking advantage of the I/O bandwidth of each
> of them. It's not really intended simply for sharing data between nodes,
> but for applications running in parallel on multiple nodes, with high I/O
> requirements. It's also a separate, chargeable product.
>
GPFS also works with multinode attached storage. We have a 3 node
regatta system, where the disks are connected in loops going by all
nodes. The GPFS/HACMP solution will then communicate all metadata over
a dedicated network (minimum 100mbit), and all real data goes
directly to disks via the SSA loops.
-jf
This archive was generated by hypermail 2.1.7 : Wed Apr 09 2008 - 22:16:15 EDT