[Beowulf] Re: vectors vs. loops

Philippe Blaise philippe.blaise at cea.fr
Mon May 9 02:56:04 PDT 2005


Robert G. Brown wrote:

>On Wed, 4 May 2005, Philippe Blaise wrote:
>
>  
>
>>>Just as any market place it is ruled by price/performance. One could argue 
>>>that during the "vector machine" epoch there was an artificial market that 
>>>was due to the cold war. Performance at any price was an important 
>>>strategic issue.
>>>
>>> 
>>>
>>>      
>>>
>>You really think, that even today, the market place is ruled by price / 
>>performance ratio ?
>>So why don't we use millions of slow and cheap cpus ?
>>With this kind of argument we must stop the research programs, and try 
>>to transform all the
>>scientific codes for embarrisingly parallel computing !
>>    
>>
>
>Um, we DO use millions of slow and cheap cpus.  And most of the
>important scientific codes have been transformed for embarrassingly
>parallel computing -- it is what "Grids" are all about.  The ones that
>cannot be done EP up front have in many cases been parallelized the hard
>way to run on network-based cluster architectures.
>
>This is the beowulf list, remember?  That's what it's all about!
>  
>
Sorry, I was on holidays.

"What makes a Beowulf cluster ?"  -> 
http://www.beowulf.org/overview/index.html
*"....Class I clusters* are built entirely using commodity hardware and 
software using standard technology such as SCSI, Ethernet, and IDE. They 
are typically less expensive than *Class II clusters* which may use 
specialized hardware to achieve higher performance...."

If all the subjects on this mailing list were about the grid and 
embarrassingly parallel computing (class I cluster),
of course I will be wrong.
But this is not because the (cluster) market is presently "dominated" by 
grid gigabit clusters that we must forget the
class II clusters.
In fact, it's not the case, and there are a lot of threads about HPC on 
this mailing list.

>And the reason the list STILL exists, some ten years after its
>inception, and is as active today as it ever has been (which is so
>active that it is a serious drain on the participants time) is because
>the marketplace is ruled by price/performance, and cluster-based HPC is
>a price/performance winner compared to nearly any alternative by a
>factor of anywhere between 3 and 100 for "most" (that danged word
>again:-) HPC applications.
>
Well, you talk again about HPC applications....

Why the costly alternatives still exist ? because the price ratio is 
more or less equal to 3,
and because high performance is a strategic issue (even in a grid 
context), and even for
a small (not conservative) lab.

>  
>
>
>Clusters, on the other hand, have proven SO successful that they HAVE
>taken over the world except for small, protected, costly but still
>powerful island protectorates.  Clusters are everyman's supercomputer --
>you can have one at home (I do:-).  They are "democratic" where big-iron
>is still a monarchy and requires a sort of vendor-allegiance and ability
>to pay high prices.
>
>  
>
Please, that's where I think you are wrong.
The adoption of cluster technology is not so simple for many of us,  
(except for the guy who has to pay the bill) :
.we need good hardware solutions to be competitive : the price / 
performance is not the only factor,
.when you decide to use a linux cluster, software environment and 
available tools are not so well designed
for high performance applications.


>As for the marketplace at any given instant meaning nothing:  Wow man,
>that's kind of Zen don't you think?  I mean, what DOES anything mean,
>after all, especially in a snapshot taken at any given instant?
>
>  
>
The actual market place seems to say:
"buy big beowulf/linux clusters for the big embarrassingly parallel 
programs".
and some translates this by: "buy big linux clusters for the parallel 
programs".
It's simplistic and dangerous.
In ten years, I'm sure that there will still be some kind of vector 
machines.
And if the criterion is only the price / performance ratio in a grid 
context, be sure that
the next generation of clusters will be under ms windows with intel cpus 
; (stupid users will
have no more to do than a clic on the left mouse button).

Phil.




More information about the Beowulf mailing list