[Beowulf] A simple cluster

[Joe Landman]

Peter St. John wrote:
> I recall discussion of the hybrid approach, which I think most of the 
> list doesn't much like, but interested me on account of my application. 
> But I hadn't realized that hybrid was required by OpenMP for multi node 

OpenMP is shared memory.  You can "fake" shared memory using a variety 
of techniques (Intel Cluster OpenMP does some of this).  You can create 
a real distributed shared memory using ScaleMP tools.  But at the end of 
the day, the OpenMP spec assumes that each thread has equal access to 
all of the (global) data.  Yes, there are thread private bits, and with 
some work you may be able to get the access to global data with some 
sort of abstraction layer (e.g. not-cheap-in-time).

> architectures. So yeah, I'll just go with MPI for starters. When I start :-)
> Peter

MPI isn't perfect, but mixing multiple models together doesn't make it 
any easier to program (somewhat harder).

It might be worth noting that the advancing of many-cores are upon us, 
so that all your nodes will look like "large" SMPs with a memory 
hierarchy and a bandwidth wall.   With accelerators they will look like 
asymmetric multi processors (aSMP), often with differing ABIs.  Lots of 
ways you can program those as well.

We live in interesting times :)

> 
> On 6/26/08, *Geoff Jacobs* <gdjacobs at gmail.com 
> <mailto:gdjacobs at gmail.com>> wrote:
> 
>     Peter St. John wrote:
>      > Geoff,
>      > Oops! I totally misunderstood it. So it's strictly shared-memory, and
>      > requires something like MPI for crossing nodes. Gotcha. Big
>     mistake, thanks.
>      > Peter
> 
> 
>     Shared memory only, yes. Many, many people skip OpenMP completely and go
>     pure MPI. From a coding standpoint it's far easier to multiprocess using
>     one technique rather than two, and the performance gains for using both
>     tend to be marginal or non-existent -- at least in my experience.
> 
>     There was a long discussion a while back on the list about the pros and
>     cons of each approach.
> 
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf


-- 
Joseph Landman, Ph.D
Founder and CEO
Scalable Informatics LLC,
email: landman at scalableinformatics.com
web  : http://www.scalableinformatics.com
        http://jackrabbit.scalableinformatics.com
phone: +1 734 786 8423
fax  : +1 866 888 3112
cell : +1 734 612 4615