Windows HPC

[Mark Hahn]

> In response to several questions/comments posted recently, here's a pointer
> to a Cornell Institute for Social and Economic Research (CISER) case study
> on how they use parallel SAS on Windows-based clusters.  

summary of the pdf: they auto-installed, so admin costs were low.
apps had special parallel support already, so they didn't do anything custom.
in short, it could all be done on Linux, almost certainly faster,
and obviously much cheaper.  but for sites who have only windows admin
skills, and who have already-adapted apps, it's a viable option.

> I think it's telling as a straight forward upgrade for users of the SAS
> statistical packages.  Something that easily and fairly transparently
> allows folks like sociologists and economists who are big statisticians to

I would argue that they are small statsticians; do you have any numbers
to back up that claim?  for "big" I'd consider 100 CPUs for 2 months
using 100 GB working set, pulled from 2 TB backing data.  merely seeding
a horde of tiny montecarlo sims from a gigabyte database is definitely
not big.

> exploit the power and scalability of Windows-based clusters. 

could you define what you mean by "power" here?  I don't see Cornell
doing anything which cannot easily be done on alternative setups,
so they have equivalent power.

"scalability", I understand.  so in what way are Windows-based clusters
more scalable?  a correct answer must define which resources limit the 
scaling of non-Windows clusters, and how Windows clusters avoid that limit.

power and scalability are well-defined technical terms, not buzzwords
that marketing types can just emit around at random.

> Yes, Windows
> clusters can run traditional MPI and PVM applications, but also commercial
> applications like this that have native support running on Windows
> clusters.   

I'm not sure anything in that paper counts as "native";
yes, SAS and Cornell did some work to adapt their code to Windows 
clusters, but I see no signs of Microsoft contributing anything
other than standard infrastructure like sockets and DNS.