[rescue] More bits and pieces

[Francisco Javier Mesa-Martinez]

On Wed, 11 Jun 2003, Dave McGuire wrote:

>    Sure, but Francisco, man...you say that about EVERY
> architecture/machine.

Not really, I only despise a few... then again I am getting a PhD in
architecture so I get to read tons of papers on failed systems, so I get
to be picky about it. As far as I know I have expressed my issues with the
R8K and MasPar (unless I am forgetting someother rant) so it is hardly ANY
architecture. I am afraid I have to eat, drink and sleep architecture for
a living (or lack of) so I am very opinionated. Specially after having to
run TONS of benchmarks and programming machines at low level, and spend
hours going *Why the F* did they do that?*

The problem with MasPar was that even the most simple instruction would
take a horrendous number of cycles. And they tried to put an interesting
routing mesh on top of that, so it was a good idea that got killed in
performance because even the great throughput associated with SIMD
machines could not offset the ridiculous latencies. Hence MasPar went the
way of the dodo..... They had a great development environment though, and
their compiler together with C* from ThinkingMAchines had some great ideas
on how to support data parallelism at high level of abstraction.

Ours has been power off for way too long, it is almost not worth the
electricity consumption.


> > We do lots of SIMD research over here, which is nice for application
> > specific tasks... it is not great however to run an OS on them. In fact
> > the MasPar acted more like a co-processor to the frontend machine,
> > than a
> > general purpose machine.
>
>    ...as supercomputers SHOULD be.

I was just pointing that out because someone was wondering about having a
sIMD machine as a general purpose machine, which it shouldn't be.

I agree with you there.

So yeah I am opinionated, however you are too so we are even. :)