Thoughts
As I prepared to deliver the ZMAXdp back to Armari, I had no idea that inside their building sat two pieces of supercomputing history. My Editor-in-Chief, David, said, "ask Dan about his Crays while you're there". And thus after a little persuasion (none really, Dan loves the Crays and would talk all day about them if he could) I got the sneaky peek and some pictures to supplement the smaller piece on what Armari is like on the inside.His wife has barred him from collecting another and that suits him I think, since he's got one of each type that he wanted to collect; MPP with proprietary vector processors and MPP with off-the-shelf CPUs. Both are liquid cooled, adding to the mystique, the T3D is unit number 1 and the T90, well that's a T90, it speaks for itself.
I'd have done a bit on the Origin 2000, but it's only got 16 CPUs, it's air cooled and it actually works! Unless it's liquid cooled, has more than 16 CPUs and has been decommissioned for security reasons (you can't have working supercomputers being bought by people that want to do evil things now, can you?), I'm not interested. Just kidding, I'll revisit the Origin in the future for a piece on the origins of NUMA, since the O2K was SGI's first ccNUMA-based computer with each CPU connected to all others. AMD's Opteron is a NUMA architecture and as we cover Opteron in further depth as time goes by (no doubt using Armari servers!), the bit on NUMA will appear.
Hopefully a peek at a couple of computing legends tickles your geekbuds. I know it tickled mine.
"Over the past fifteen years or so, people have relied upon me to design and integrate them a few high performance computers. During that time I've only ever met five people I'd wholly trust to do the same for me. Those five people include myself and Dan Goldsmith." Paul Dutton, HEXUS.net
Corrections
A short while after the article was published, someone got in touch to correct my assertion that Sun's original Starfire MPP system was a Seymour design. So while Cray did sell the design to Sun, Seymour himself had no real input into the system. They acquired the design from a company called Floating Point Systems when they purchased FPS after FPS had some difficulties. Renaming FPS's FPS-500 created the Cray S-MP and Clark Masters, FPS employee and now VP and General Manager of Enterprise System Products at Sun, continued to oversee that project.Subsequently, following Cray's purchase by SGI, the S-MP division was sold to Sun. That then became the Starfire. So a Cray product at one point, but with its beginnings at FPS and under the masterful eye of Clark.
Many thanks to Robert E. King, a systems engineer at Northrop-Grumman, for the correction. He worked on the Starfire's ancestors and has intimate knowledge of the product in its early days. Here's his email to us, verbatim. Cheers Bob!
As a systems engineer, I had to supervise the installation and operation of four of the ancestors to the Sun Starfire servers. I thought I would correct the errors in your article at http://www.hexus.net/content/reviews/review.php?dXJsX3Jldmlld19JRD05MDMmdXJsX3BhZ2U9MQ==
I had many fascinating opportunities over a decade to discuss the on-going design effort with the real designer of the Starfire servers.
Seymore Cray had NOTHING to do with the design of the Starfire or its ancestors!
In the 1980s, a company named Floating Point Systems in Beaverton, Oregon was a manufacturer of peripheral array processing units (vector processors) that could connect to the Digital Equipment Corporation series of VAX computers. In the late '80s, Clark Masters started a design for FPS's own first complete computing system. That machine was called the FPS-500 and had an innovative architecture. A single or dual CPU (scalar unit) was combined with one vector processing unit per CPU and an optional matrix processing unit. The original CPU design was a home-grown effort, later replaced by a high-speed ECL implementation of the Sun SPARCstation2 processor. The service computer subsystem that provided diagnostics and system management features was based on a Sun IPX.
When FPS began having difficulties, Cray stepped up and bought FPS. They quickly renamed the FPS-500 as the Cray S-MP (for symmetric multi-processor). The system still used the ECL implemetation of a SPARC CPU. Clark Masters was still in charge of the S-MP devision.
Later Cray was purchased by SGI and the S-MP division was sold to Sun Microsystems. Sun had the financial resources to allow Clark to implement the latent features of the design as the original Starfire 10000. Since then, he has expanded the design to include the 12K, 15K, and the new 20K systems.
Robert E. King
Systems/Networks Engineer
Northrop-Grumman/ TASC, Inc
