[Bioclusters] cluster hardware question

[William Park]

On Thu, Jan 16, 2003 at 11:41:23AM -0500, Jeff Layton wrote:
> > > If you assume that you are running on one instance of your app per
> > > CPU and that you can only address 3.5 Gig of memory per CPU, then
> > > that leaves you with around 8 Gigs to play with (giving a generous
> > > 1 Gig for the OS). While RAM is expensive compared to disk, this
> > > idea is also much faster than disk. Would 8 Gigs be enough for some,
> > > many, lots of people?
> >
> > Some.  I think the PAE stuff lets you address up to 64 GB, but I dont
> > know how well it works.  I have heard it slows things down a bit.  Using
> > it as a ram disk for swap space (ala the old Crays with the SSD) could
> > be interesting.
> 
> We're toying with that idea as well. Also, we're looking at using the
> extra memory for a diskless node, so you can copy what you need
> to the RAM-disk. This way the nodes can be moved around from
> one secure area to another without having to swap disks.

Uhh?  You are aware that 100Mbit is about 10MByte/s which is about the
same as sustained rate of IDE disks.

-- 
William Park, Open Geometry Consulting, <opengeometry@yahoo.ca>
Linux solution for data management and processing.