[rescue] SMP on intel wasteful?

[Francisco Javier Mesa-Martinez]

On Wed, 26 Jun 2002, Dave McGuire wrote:

> On June 25, Francisco Javier Mesa-Martinez wrote:
> > The truth is that discussions like these are moot. Vector is hardly used
> > anymore,
> 
>   Umm, what?  Do you not have much contact with the research community
> or something?

I have a LOT of academic contact sire, way more that thou so I do not have
anything to prove in this aspect.

  There's a LOT of vector stuff going on in areas of
> fluid mechnics and oil exploration, since (apparently) both of those
> areas have a few problems that map very well to vector machines.

A lot of the stuff that was used on vector machines was ported to normal
SMPs. Specially oil stuff, they were using finite element analysis anyway
(i.e. grid based subdivisions) that are also "easily" portable, everything
is relative of course, to scalar SMPs. Saying that there is a  LOT of
vector stuff is a misstatement (and I love vector machines, BTW) case in
point the SV2. This machine is almost 100% subsidiced by the US gov..
there is just no academic/commercial demand for it right now. Then it is
also a matter of culture, Japanese have been more successful in
building/marketing Vector supers in the past 2 decades or so. And the US
has concentrated on SMP approaches, but the Japanese could not sell the
machines in the US... so a lot of the vector groups moced to SMPs. 
 
> > and when it is usually specialized SIMD approaches are used in that
> > field.
> 
>   Umm, WHAT?  Show me one vector machine that ISN'T a "specialized SIMD
> approach".

This is a fallacius statement. A little bit of logic will help here...
Vector is a subset of SIMD, therefore not all SIMD approaches are Vector.
Thus my point some of the stuff that used vector processors, was moved to
more specialised SIMD designs (which btw is my are of research) like
systolic arrays, 2D flow grids, et al.

 > 
> > Granted each approach has its niche, clusters are better for
> > non coupled data problems, whereas vector/SIMD machines are better for
> > data parallel problems. Once you have specified your problem, then by al
> > means chose the RIGHT tool to solve it, And let's stop this pissing
> > contest, please.
> 
>   Each approach has its niche, sure...and both approaches ARE niches.
> If they weren't, there'd be vector machines and clusters everywhere.
> 
>   The fact that big Linux clusters get a lot of press does not translate
> to "vector is is hardly used anymore".

They have clearly displaced the spotlight from the traditional vector
super. In some sense they have democraticied the playing field. The fact
of the matter is that more and more code is moving away from vector
machines than towards vector supers. All I have to do is to look were most
of the money is going for research from NSF/DARPA.... or attend any HPC
conference. This doesn't mean that I like it, but that is the reality. The
writing is on the wall, and hadn't it been for the US orders CRAY, the
only American vector company left, would be no more. Heck, evey CRAY of
all people is reselling clusters of DELL FeeCees! But there is
hope I guess, the Earth simulator is a big onking cluster of Vector
machines.... but then again it is not an American machine.