[geeks] Drool-worthy

[Francisco Javier Mesa-Martinez]

On Wed, 10 Nov 2004, Shawn Wallbridge wrote:

> >
> > * It implies that the Origin2000 was one of the first NUMA systems. I
> >   don't believe it was ... Sequent and similar was far earlier. The
> >   ccNUMA architecture did address some weaknesses of the NUMA
> >   architecture.
>
> He said the Origin 2000 was _SGI's_ first NUMA machine.

Actually NUMA-Q and CC-NUMA came out at pretty much the same time, in fact
I believe CC-NUMA was ready on late 95, and NUMA-Q did not came out 'til
late 96. But they are pretty much contemporary. The CC-NUMA was a direct
descendant of the NUMA project at stanford so it may predate the NUMA-Q
even further, it was also more scalable. NUMA-Q however used intel based
processors, so that was a first I assume. DG and Unisys did release NUMA
machines much later...

> >
> > * The AMD Opteron is a NUMA system ? I suppose it might have on chip
> >   support for some NUMA architecture, but surely NUMA is more the
> > system
> >   than just the CPU ?
>
> I don't know what he was talking about here.

The opteron implements hypertransport and the memory controller on chip,
Hypertransport can be used to build NUMA configurations pretty easily, you
can get and 8 processor NUMA node with almost no glue logic. More
processor configurations will need hypertransport bridges across nodes
though. But any Opteron with more than 2 processors is in fact a NUMA
machine. The memory is distributed accross processors... each processor
accessing directly one or more banks of memory, and having to go through
their neighbouring hypertransport channels for memory requests not found
in their local bank.

> > * It mentions the StarFire as being designed by Cray. Wasn't it just
> >   the CS6400(?) designed by Cray ? Admittedly the E10K is quite
> >   different to the other USII Enterprise line.
>
> IIRC Sun bought part of Cray before or during the SGI purchase, so he
> could be kind of right. I would have to look into it more. Either way,
> the starfire name didn't start until the 15k's (IIRC), so that would
> have been long after Cray was gone. He may be thinking of the E10k.

The E10K was supposed to be based on the CS6400 architecture that SUN
bought from CRAY when it was being swallowed by SGI. Most of the E10K
designers came from that CRAY group.

SGI did not seem to have gotten much out of CRAY, in fact the "CRAY-Links"
have nothing to dod with CRAY technology oddly enough...