Re: [PLUG] super computing/parallel programming inquiry

["Gavin W. Burris" <bug@sas.upenn.edu>]

Mark speaks truth.

GPGPU is only for certain classes of problems and is a huge investment
in developer time.  Also, I don't consider sticking one, or multiple,
GPU cards in compute nodes an insignificant cost.  Proper GPU cards can
double your hardware cost per node.  They will also eat a lot of
electricity.

I have built many hpc clusters, based on CentOS.  ROCKS is CentOS with a
lot of consideration for the cluster deployment.  I did a series of blog
posts on a recent hpc cluster deployment:
http://idolinux.blogspot.com/2010/06/hpc-cluster-install-infiniband.html

This is a topic I'm highly interested in.  Let us know what you decide on.

Cheers.


On 03/03/2011 05:12 PM, bergman@merctech.com wrote:
> The pithy ruminations from Doug Stewart <zamoose@gmail.com> on "Re: [PLUG] super computing/parallel programming inquiry" were:
> 
> => On Thu, Mar 3, 2011 at 4:48 PM, Mike Sheinberg <m.sheiny@gmail.com> wrote:
> => > I'm doing some research into viable super-computing architectures for a
> => > software project and was hoping someone could point me in a good direction
> => > based on experiences. Specifically I'm looking at the two platforms at the
> => > moment (but open to other suggestions to look into):
> => =
> => 
> => Have you looked into Rocks and/or Rocks+?
> 
> ROCKS is a nice cluster management framework. It has some implications about the type of computing, but doesn't restrict (or require) MPI, OpenMPI, GPU, or other techniques.
> 
> I run a ROCKS cluster (286 cores now) and am working on building a new 350+ core cluster (then merging the two). 
> 
> [Dammit Jim] I'm a sysadmin, not a programmer[1]...but it sounded to me like Mike Sheinberg's question about GPU vrs MPI really, really, really needs to be addressed by the developers or application architect, not a system administrator. Either technology may appropriate for some types of computation, but they require very different programming approaches. This is a case where I'd let the application needs, the ability/interest of the developers, the intended market, the production constraints (ie., time to produce results, scalability), and budget be the main drivers of the technology choice, not the cluster management framework.
> 
> I'd be happy to contribute my scant knowledge about ROCKS and our cluster, and may be able to put Mike in touch with people who can help direct the technology choice to meet the application needs...Feel free to contact me off the list.
> 
> [1] http://knowyourmeme.com/memes/dammit-jim-im-a-doctor-not-a-x
> 
> Mark
> 
> => 
> => http://www.rocksclusters.org/wordpress/
> => http://www.clustercorp.com/index.html
> => 
> => LJ had a decent article about it a while back:
> => 
> => http://www.linuxjournal.com/magazine/building-linux-based-high-performance-compute-cluster
> => 
> => 
> ___________________________________________________________________________
> Philadelphia Linux Users Group         --        http://www.phillylinux.org
> Announcements - http://lists.phillylinux.org/mailman/listinfo/plug-announce
> General Discussion  --   http://lists.phillylinux.org/mailman/listinfo/plug
> 

-- 
Gavin W. Burris
Senior Systems Programmer
Information Security and Unix Systems
School of Arts and Sciences
University of Pennsylvania
___________________________________________________________________________
Philadelphia Linux Users Group         --        http://www.phillylinux.org
Announcements - http://lists.phillylinux.org/mailman/listinfo/plug-announce
General Discussion  --   http://lists.phillylinux.org/mailman/listinfo/plug