[Beowulf] Windows cluster

[Ryan Sweet]

I have had a customer that desired a serious evaluation of Windows as an HPC 
platform, because they had an in-house application that had been written on 
Windows.  We came up with a feasible design, that basically reduced the 
administrative headaches etc to a minimum, and used a Linux box as the head 
node with a web portal for job submission, etc, but still advised the customer 
that it would probably cost them less to port the (relatively small) 
application to Linux.  Our design leveraged a lot of customisation of Windows 
at a fairly low level, and while it was likely to be quite easy to maintain, 
it was not suitable for implementation by an average Windows admin.

In the end the customer agreed and decided it made more sense to port the 
application to Linux.

I would add my voice to the others to say that even if you don not have an 
experienced Linux admin, it will probably be easier for you to build and run a 
Linux cluster alongside your windows domain than to build an actually 
efficient Windows based HPC system.

However, if you are stuck with Win32, there is a Microsoft initiative centred 
around HPC, which is in beta:

http://www.microsoft.com/windowsserver2003/hpc/default.mspx#EDBAAA

It includes and SDK, MPI, and a resource manager, though I have not tried them 
directly.  Microsoft was giving demos at ISC2005, and the reps there hinted 
about providing significant assitance for early adopters.  If you are serious 
about running a cluster with a Win32 OS, you should probably check it out.

Keep in mind that as with Linux HPC, you will want to strip Windows to the 
bare minimum services necessary to run the OS, your resource manager, and your 
applications.

Windows by default starts a huge heap of services that do not make sense at 
all in a pure HPC environment.

Make sure the thing is firewalled (not just nat'd) in both 
direction, and insist that you maintain tight control over which executables 
you allow to be run on the cluster.

Regarding MPI, in addition to the commercial MPIs, LAM-MPI should build on 
Cygwin, though I'm not sure what sort of performance you could expect in 
comparison to Linux on the same hardware.

regards,
-Ryan

-- 
Ryan Sweet             <ryan.sweet at aoes.com>
Advanced Operations and Engineering Services
AOES Group BV            http://www.aoes.com
Phone +31(0)71 5795521  Fax +31(0)71572 1277