[Beowulf] Re: Q&A: BeOS and IBM Cell processor

[Vincent Diepeveen]

----- Original Message ----- 
From: "Ed Karns" <edkarns at firewirestuff.com>
To: <beowulf at beowulf.org>
Sent: Monday, January 23, 2006 5:51 PM
Subject: [Beowulf] Re: Q&A: BeOS and IBM Cell processor


> q1) How about the BeOS? ... any reports on attempts or successes in  using 
> BeOS on clusters and multiprocessor systems?

For a good operating system there is 2 conditions IMHO to take off:

  a) good driver support
  b) good compiler that generates faster executables than competitors

A is trivial.
B is less trivial to most, but consider this. I bought a dual core opteron 
2.4Ghz,
under windows visual c++ 2005 professional edition is 10% faster as it seems 
now
than any compiler is under *nix.

That means that for the same money i paid for this box, under linux i need 
to
buy a dual core opteron 2.64Ghz.

Now 10% is still not much. I do know there is colleges of mine where it is
actually 60% speedwin under windows.

This is a hard undeniable truth, some people are fulltime busy overclocking 
their
system 1 or 2%, having 60% speedwin is a lot then.

So the OS is not relevant there, *compiler* is.

GCC always is the slowest of course, unless there is political dogfights 
getting fought out,
so you need a compiler faster than GCC.

> Considering that the latest iterations of BeOS seems to have been 
> specifically targeted at multiprocessor, multitasking projects, this 
> might seem to be a natural ... and not just a toy of garage shops

By now even windows is 64 bits, so unlike they have more drivers for it than 
microsoft has for 64 bits,
it'll be hard to use beos for the average user.

> q2) Anyone examining the IBM Cell processor? ... Is this processor  worthy 
> of consideration for clusters and multiprocessor systems?

No matter how impressive a new cpu looks like, what matters in the end is 
the date and price for which average users can build themselves a system 
with it.

On paper it's a most impressive cpu. 8 help processors which would clock at 
4Ghz. Mass production of the cpu for
some playstation type machine, keeping the price a cpu very low.

A cheap processor which has 8 cores running at 4Ghz, capable of doing a 
number of multiplies a cycle is very interesting for the highend
of course in order to cluster it for embarrassingly parallel software such 
as matrix multiplications.

If you can buy such a system for a 1000 euro a box and another 2000 euro a 
box for network and you fill a sporthall with it,
then another professor has his few days of glory at top500.org.

It is very hard to write software for the cell processor. You need top 
programmers that really know how to parallellize software at cycle
level. Each help-processor has to ship and receive blocks of memory over the 
central bus, this within a 256KB cache space. That's all
very hard to program for.

So from highend viewpoint it *can* be a great processor, provided you have 
professional libraries that ease the task of researchers who
want to run a specific application.

It is an ugly processor if your software has branches and/or isn't optimized 
to this specific architecture.

> Same, same for the IBM Cell processor = targeted at the future 
> multiprocessor markets and not just game systems (like Sony  Playstation 
> 3).

Please don't compare highend software, which usually is crap software 
because scientists prefer to write software once and then
the coming 30 years run with it until they retire and usually very unstable 
software, just bugfixed barely so it works for their own
machine. Just don't compare that with game systems that run at millions of 
machines, where it is important to .

In general game software is the most optimized and most well written 
software on the planet.

My own engine code qualifies for that too and our 3d engine is very 
optimized too by now (though not at cycle level yet).
Things really get figured out at cycle level in general.

The average person hugely underestimates how well optimized mass software 
is.

So that the game programmers get a cpu in their hands that's probably 
amongst the most difficult to optimize for,
is not a real surprise to me. They just need something that has a real good 
bestcase if you know how to program for it.

So that PS3 will just destroy everything in terms of game performance, you 
bet all those game programmers will fly on that chip.
In fact the chip delivers a lot more than what you need for games.

For those who are doing matrix multiplications this chip potentially crushes 
everything, especially because it should have a low price a chip,
as Sony should sell a million or 100 consoles with that chip inside.

That *garantuees* a low price for the chip. How to do the network for it is 
an entire other question, let's leave that to the networking
experts to solve.

You'll see that IBM wants to earn too much on this chip. If they wait too 
long selling it for a low price, of course AMD and Intel and perhaps other 
manufacturers will have released by then some other chip that's faster.

Vincent

> Ed Karns
> FireWireStuff.com
>
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit 
> http://www.beowulf.org/mailman/listinfo/beowulf
>