Re: MISC-d Digest V99 #87
Jeff Fox writes:
> Interesting. But who said that dsp-clusters are off topic. I know
Thanks.
> that DSP have large die and such but I don't see why a MISC cannot
> be a DSP variant or that a DSP chip has to have more than a minimal
> instruction set.
The problem is that this is not about hardware design: there are
pretty useful *affordable* chips out there. Furthermore, while I
intend to go the Forth way, I don't want to keep the C people out. The
more people participate, the better.
> I can grasp the concept of an OpenSource paradigm. A lot of Forth
> projects have been of that nature. Beowulf paradigm I don't quite
> get. Maybe I am missing something. I think of Beowulf as just a
> networked PC with software layers for multiprocesing on that platform.
Admittedly, it's no big deal, but the name "Beowulf" has become a
trademark/meme cluster. Whenever you say "I'd like to bring the
Beowulf to DSPs" more people will grok it.
> If some DSP has networking hardware one could implement the software
> layers that are used in Beowulf on the existing hardware there. I
> would expect it to reduce system performance by an order of
> magnitude rather than boost it. I just imagine that the layers
Sure, but IP stacks can be ridiculously compact (half a kiloword of
assembly), and there are tons of PVM/MPI software out there.
> in Beowolf are typcial of the environment and might not be required
> on the DSP.
Of course not, but one needs an application base first.
> If you have something that has functions of distributed shared memory
> in hardware I would think you would want to try to remove extra software
> layers to boost performance. If you model the software after stuff
Definitely. This is one of SHARC nicer features: sending messages by
executing a machine instruction vs. kernel traps and NIC register banging.
> that evolved getting network hardware designed to compete with hard
> disks to act like a multprocessor bus the software might have layers
> of complexity that you don't need if the DSPs were designed to be
> clustered in the first place.
True, but one needs to get started somewhere. It makes definite sense
to initially let a (small) Beowulf to get the hosting: as a number of
multi-DSP cards in PCI slots. If it catches on one can think about
interfacing video and EIDE devices and whatnot to the DSP.
> I must be missing something. Maybe I am just thinking of the details
> of the Beowulf implementation that are system specific too much instead
> of just the idea of a standard way using clusters.
>
> If you are talking about the multiprocessing software being Forth
> or looking like a MISC variant to a user why would that be off
> topic in the MISC mail list?
We can definitely crosspost relevant materials both here and there.
> >Discussion topics include both hardware and software issues.
> >
> >Regards,
> >
> >Eugene Leitl
>