[rescue] SGI files Ch11 again, bought by Rackable for $25M

[jodys at helluin.org]

On Wed, Apr 01, 2009 at 10:30:35AM -0700, Brian Roth wrote:
> The manual has the NUMA cabling defined for different sized systems. Using Craylink and expresslinks and I bet the HIPPI was used between the Meta Routers.
> 
> http://techpubs.sgi.com/library/tpl/cgi-bin/download.cgi?coll=hdwr&db=bks&docnumber=007-4192-001

HIPPI was a lot slower though, 800Mbit in it's first revision, compared
with 800MB for NUMAlink2.

I'm pretty sure there were different part numbers for different length
cables and there were restrictions on what device could use what cable. 
The XIO/Crosstown links used the same cables, but were a different system.

I found a pdf[0] on Ian Mapelson's excellent site describing the system in
question. There were qty (48) 128p machines connected with 12 HIPPI ports per
machine, connected via a 3-dimensional toroidal interconnect using 36 
HiPPI-800 16 port switches. It appears that they upgraded to HIPPI-6400
in 1999. It does not appear that HIPPI was used to connect the meta 
routers, which would be required for 128p machines. Some sort of 
standard job control system was used to distribute jobs among the 
48 different machines, it appears to be a cluster of really big 
computers using HIPPI as the cluster interconnect, not as part of the 
NUMAlink on the SGIs.

[0]http://www.futuretech.blinkenlights.nl/origin/ASCI_fly.pdf

Jody

> 
> --- On Wed, 4/1/09, jodys at helluin.org <jodys at helluin.org> wrote:
> From: jodys at helluin.org <jodys at helluin.org>
> Subject: Re: [rescue] SGI files Ch11 again, bought by Rackable for $25M
> To: "The Rescue List" <rescue at sunhelp.org>
> Date: Wednesday, April 1, 2009, 1:27 PM
> 
> On Wed, Apr 01, 2009 at 11:18:17AM -0600, Robert Darlington wrote:
> > HIPPI was the world's first gigabit networking system.  Friends of
> mine
> > worked with SGI to develop it, got an R&D 100 award out of the deal
> etc.
> > Major breakthrough at the time (late 1990s I guess, before my time there).
> > When I came on, I was mostly handling software and OS stuff so I
> didn't have
> > my hands inside the machines too often.  I do remember seeing the thick
> > cables though and having to work around them.  One "router rack"
> with 4
> > O2000 racks on each side with the heavy HIPPI cable interconnects.  So 9
> > racks total per row, times 48 rows deep.   LOTS of cable.
> > 
> > Later we experimented with the Origin 3200 line, testing R bricks (routers
> > for the ccNUMA interconnects) vs Myranet and such.   Good times.
> 
> Wow! Origin 2000 was 16 CPU per rack, so 9 racks times 16 is 144, 144
> cpus times 48 rows is 6912 cpus. No wonder they used HIPPI, I believe that
> Origin 2000, with the last routers, maxed out at 512 cpus. 
> 
> Just to clarify, HIPPI and NUMAlink are two seperate things. I imagine
> that NUMAlink connected up Origin 2000 compute nodes into larger 
> cc-NUMA systems, those larger systems then must have been interconnected 
> with HIPPI. 
> 
> I have some fiber HIPPI gear for my SGI gear and, one of these days, I'll
> get 
> around to setting it up.
> 
> Jody
> _______________________________________________
> rescue list - http://www.sunhelp.org/mailman/listinfo/rescue
> _______________________________________________
> rescue list - http://www.sunhelp.org/mailman/listinfo/rescue