Re: swap

[Douglas Allan Tutty <dtutty@porchlight.ca>]

On Thu, May 10, 2007 at 08:14:11PM -0500, Ron Johnson wrote:
> On 05/10/07 18:10, Douglas Allan Tutty wrote:
> [snip]
> > 
 
> And if you say "make a SSD", then I say that it's still slower than
> RAM because disk channel bandwidth is still *much* slower than
> memory bandwidth.  Better to spend that money on extra RAM.
> 

But if the problem/program can only run on one processor, and you have
maxxed out the memory that that processor (or the MB it's on) can
handle...

I've been trying to find (e.g. google, ibm) information on
supercomputers/HPC that are built to solve sequential problems and
striking out.  Everything is on paralellization and how clusters don't
help if you can't make it a parallel problem.  

The other tack I've been thinking is what if you could make a virtual
computer that ran on more than one CPU/node that was _more_ powerful
(compute, memory, whatever) than any one CPU/node.  This would be the
direct opposite of the current virtualization of a guest being a subset
of the host (e.g. Xen, z/VM).  The problem here is that, unlike Xen, it
would require a true emulation rather than passing code directly to a
processor.  

Other than your bach job, I wonder what would be an example of a current
real-world sequential compute problem; a long-running program that
couldn't be run in parallel on multiple nodes.

Doug.