[Beowulf] What's the category of Beowulf among Clusters?

Geoff Jacobs gdjacobs at gmail.com
Fri Dec 19 17:58:14 PST 2008


Mark Hahn wrote:
>>     Clusters, Grids, MPPs, MPI, OpenMP, HA, LB, GPGPU, FPGA, SMP, NUMA,
>> SSE etc..
>>     These abbreviations and terms almost cram my head, so I have to
>> redvelop and re-index them in my memory(brain).
> 
> think of what the acronym is abbreviating, and the logic of that name.
> 
>>     As a newbie, when I read the articles in wikipekia, I got confused.
>>     In the segment Cluster categorizations
>> <http://en.wikipedia.org/wiki/Computer_cluster#Cluster_categorizations>
> 
> that's horrible and incomplete.

It's wiki. Why don't we fix it?

> IMO, HA and load-balancing are not really distinctly different, since LB
> is really just active-active HA.  (or HA is active-passive LB).
> other than HA/LB, clusters are computational.  within compute clusters,
> the main distinction is how tightly coupled they (or the programs they run)
> are.  grids are the extreme loose end: basically no inter-node
> communication,
> often geographically distributed, often ad-hoc collections of different
> kinds
> of machines run by different organizations.  the opposite is a homogenous,
> tightly-coupled cluster with a dedicated local network optimized for
> inter-node communication and running few multi-node jobs.
> 
>> , both "Cluster computing" and "Grid computing" are the subclasses of
> 
> "cluster computing" is descriptive: the entity is a set of nodes somehow
> combined, usually by a local communication fabric.  by definition, the
> nodes
> are separate, so distributed.  (the 'distributed' here means that
> communication is by explicit message passing; the opposite is
> shared-memory,
> where communication is implicit and done by read/write operations to
> memory.)
> 
> "grid" is a marketing term for "loosely coupled distributed clustering";
> it was a trendy word 10 years ago, but has fallen into disuse because
> it's so generic (and not all that widely applicable).

I always thought the idea was to charge for computing as a service (just
like the electrical utility). Actually, many firms are doing this now.
Amazon, for example.

>> "Distribute computing" ,and the third one is "Massive parallel
>> processing ". IMHO, the latter category is more reasonable(right or
>> not?) .
> 
> MPP doesn't mean much; its best to avoid the term and stick to more
> specific ones.
> 
>>     However, since there are too many cluster software products, how
>> can I categorize Beowulf like clusters( loosely coupled, use MPI)? or
> 
> beowulf certainly does not imply loose coupling (or rule out PVM.)

Far from it. In fact, a great deal of work goes into optimizing the
interconnect and the software payload for tightly coupled, fine
granularity workloads.

>> what's the category of Beowulf like clusters?
> 
> beowulf is compute clustering using mostly commodity hardware and mostly
> open-source software.

Designed to reduce the clock time and/or increase the maximum practical
problem size of computational problems.

-- 
Geoffrey D. Jacobs



More information about the Beowulf mailing list