[BBLISA] physicalization

[Ian Stokes-Rees]

On 6/26/10 12:15 PM, Edward Ned Harvey wrote:
>
> Apparently, physicalization is the reverse of virtualization.  The use
> of hundreds of low-power (like atom, arm, etc) processors each running
> its own OS, instead of a smaller number of higher-power (like xeon,
> etc) processors ... To build-out highly parallelized servers, for the
> purpose of either eliminating the need for virtualization, or to run
> highly parallelized and distributed work loads supposedly cheaper and
> less power hungry than the equivalent high-power "standard" server
> solution...
>
>  
>
> I am aware of products from seamicro and rackable (sgi).  Does anybody
> have any experience in this area?  Know any other options I should
> consider?
>
>  
>
> The specific requirement we're solving is a workload which is highly
> parallelizable.  We're currently paying Amazon EC2, but finding it's
> not very economical.  So we're exploring alternatives, hoping to find
> a way to run these parallelized jobs more economically.  It is also
> ideal to keep the workload in-house, so we can eliminate the security
> concern about Amazon employees having access to our data and so forth.
>
>

Keep us updated on what you discover on this front.  People have been
trying this kind of thing in different forms for quite awhile.

We have some highly parallelizable workloads, but are in an academic
research environment and can benefit from the US Open Science Grid from
which we have (recently) been able to secure sustained access to 3000+
job slots at computing centers across the US in an "opportunistic" mode
which means we use otherwise idle compute cycles at non-commercial
research institutions (some federal, some state, some universities).

If I were you, I'd look into the 4x12 core AMD systems.  These things
are half price or less on similar Nehalem Xeon systems on a per-core
basis, but you then have to figure out if you are OK with the memory,
disk, network I/O patterns and can live with the slower AMD processors.

>From a different perspective, (and also *very* dependent on your having
a good grasp of the workload characteristics), if this is a high value
and long term workload you *may* be able to benefit from GPUs, which
effectively are hundreds of slower compute cores accessible from the
same system image, but will require a small computational kernel which
you can port to a GPU environment.

Ian
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.bblisa.org/pipermail/bblisa/attachments/20100627/a725c34d/attachment.htm 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: ijstokes.vcf
Type: text/x-vcard
Size: 380 bytes
Desc: not available
Url : http://www.bblisa.org/pipermail/bblisa/attachments/20100627/a725c34d/attachment.vcf