[Beowulf] Re: vectors vs. loops

[Douglas Eadline - ClusterWorld Magazine]

On Tue, 3 May 2005, Philippe Blaise wrote:

> Robert G. Brown wrote:
> 
> >....
> >
> >Still, the marketplace speaks for itself.  It doesn't argue, and isn't
> >legendary, it just is.
> >....
> 
> But, does the hpc marketplace have a direction ?

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.

> 
> Few years ago, some people had a "fantastic vision" to replace the 
> vector machines market :
> use big clusters of SMPs with the help of the new paradigm of hybrid 
> mpi/openmp programming.
> Then the main vendors (usa), except Cray, were very happy to sell giant 
> clusters of smp machines.
> 

The issue is economics. It cost about $5 Billion to build a FAB these 
days. It also costs quite a bit of money to design a CPU. Economics makes 
you ask questions like, "If I build a CPU how many will I have to sell 
to recoup my investment and make money?" Your answer had better be in the 
millions.   


> Nevertheless, the japanese guys built the "earth simulator" ; which is 
> still the most powerful machine in the world
> (don't trust this stupid top500 list).

And how many earth simulators have been sold to date?
(an obvious "loaded question") With proper funding very fast supercomputers 
can be built. It is really a question of price/performance. (But we all 
know the answer is "42" in any case.) 

> 
> Then Cray came back ... with vector machines...
> 
The Cray of today is not really the Cray of yesterday. Cray was purchased
by SGI in 1996 and sold to Tera in 2000. Tera then changed its name to
Cray. When in doubt check Wikipeda
http://en.wikipedia.org/wiki/Cray_Research 


> Don't underestimate the power of vector machines.
> Yes Fujitsu or NEC vector machines are still very efficient, even with 
> non contiguous memory access (!!).

I doubt anyone on this list questions the power of vector machines. It is
just that they doubt the ability of vector machines to compete on price to
performance issues. Again it is "commodity economics". Clusters leverage
the commodity market. We get great price-to-performance because companies
are making products for the things like web servers, desktops, and video
games.

> 
> One year ago, the only cpus that sometimes were able to equal vectorial 
> cpus were alpha (ev7) and itanum2 with
> big caches and / or fast memory access. Remember that alpha is dead. 
> Have a look to the itanium2 market shares.
> 
> The marketplace is not a good argument at all.
> 

Market success or failure is due to companies making money not CPU from
performce. Baring any large government programs or strategic initiatives,
the market is what will determine the future of supercomputing. Where will
the money come from?


> Vectorization and parallelization are compatible
> Hybrid mpi/openmp programming  is a harder task than mpi/vector programming.
> If you have enough money and if your program is vectorizable, buy a 
> vector machine of course.
> 

"If you have enough money" - means the market will decide. 

> Cluster of SMPs ? they will remain an efficient and low cost solution, 
> (and quite easy to be sold
> by a mass vendor).
> And thanks to cluster of SMPs with the help of linux, the HPC market is 
> now "democratic".
> 
> Of course, it would be nice to have a true vector unit on a P4 or Opteron.
> But the problem will be the memory access again.

If there were a need for such a device in the big markets then there would 
be such a device in processors. 

We must play with the "Lego" that we can find in the commodity markets.
Fortunately, there are many companies that make "specialty pieces"  
(interconnects, compilers, packaging, etc.) for the cluster market. But
like all companies they need to earn money - something the vector
supercomputers seemed to have trouble accomplishing in the absence of a
government supported market. 


Doug
----------------------------------------------------------------
Editor-in-chief                   ClusterWorld Magazine
Desk: 610.865.6061                            
Fax:  610.865.6618                www.clusterworld.com