[Beowulf] Re: SMPs + One processor machines = Heterogeneous Cluster

[Mark Kosmowski]

> Message: 8
> Date: Wed, 2 Apr 2008 21:35:34 +0800
> From: "Cally K" <kalpana0611 at gmail.com>
> Subject: [Beowulf] SMPs + One processor machines = Heterogeneous
>        Cluster
> To: beowulf at beowulf.org
> Message-ID:
>        <b05971d10804020635u6d41cfb8v115deea93a6499d6 at mail.gmail.com>
> Content-Type: text/plain; charset="iso-8859-1"
>
> Hi guys,
>
> Thanks for the previous answers to my question, did not expect that kinda
> flow of information and it proved to be helpful in my research. I have a
> different question and I hope I can get some nice feedback.
>
> I am dealing with heterogeneous cluster in my research. I am a novice when
> it comes to networking, and have only been around with cluster for about a
> year. Anyway, from my previous question, and after doing some reading, can I
> say that dual cores and quad cores are known as SMPs.?
>
> And would 1 say that, a cluster that has machines with 1 processor ( etc.
> Pentium 4 )and 2 SMPS( it can be quad or dual ) but all belonging to one
> family ( intel for example ) -- can I say that the cluster is a
> heterogeneous cluster.

My limited (and perhaps incorrect) understanding is that SMP denotes
more than one processor, regardless the number of cores that the
processor has.

Heterogenous, to me, implies a cluster of machines running different
operating systems and/or different architectures.

As an example, my personal is three Opteron systems, each of which has
two single core CPU's, all running OpenSUSE 10.2 64-bit.  This is a
homogenous cluster of SMP machines.  I plan to add a dual core, single
CPU Athlon (64) system to the cluster at some point.  This planned
machine would (in my opinion) not be SMP (although the more I think
about it, the more I think I am either wrong or in the minority camp
about SMP and cores) and the cluster would remain homogenous even if I
were to put OpenSUSE 11.0 (64 bit) on the system (the new OpenSUSE
release and my projected budget have similar timescales).

Even if I were to use an Intel Pentium system of whatever nomenclature
that is instruction compatible with my existing Opterons (the
difficulty in immediately knowing whether a Pentium chip is 64-bit
capable or not is one of the reasons I am sticking with AMD) the
system would remain homogenous.

If, however, I were to add a Sun SPARC machine as a node, now I have a
heterogenous cluster.  Likewise, if my new Athlon (64) node used
Solaris (for Intel), the cluster would be heterogenous in my opinion.

My rule of thumb for heterogeneity is whether the programs to be run
need to be recompiled for a given node.  If all of the nodes can use
one single compile of all of the codes needed, the cluster is not
heterogenous.  I guess this would mean that running a 32-bit program
(I want karma for minding my usage of code vs. program here! :) ) on a
mixture of 32-bit and 64-bit machines would be homogenous, but if on
the same cluster, the 32-bit machines ran a 32-bit program and the
64-bit machines ran a 64-bit program then that cluster might be
considered heterogenous.

I hope this at least sparks an interesting conversation,

Mark Kosmowski