by Steven Witucki, assistant editor LIVEwire
Dallas, Texas — There’s something special about the SC2000 conference this year, according to Steve Conway, Director of Corporate Communications for Cray computers. Conway believes that future conferences could mark a change in the ranking of supercomputer performance.
He’s thinking of the recent announcement from the HPC user forum organized by the International Data Corporation (IDC). The HPC user forum hopes to develop a new benchmark suite to test the overall performance of HPC systems. “What you’re really seeing behind the scenes here, I think, is a user protest having occurred,” said Conway.
HPCwire: Tell us about how the users are shaping the proposed new benchmarks.
CONWAY: The situation was that you’ve got companies in this market that are so big that this market is too small for them to customize. So the strategies that are being used are to take computers that really have been designed for lower performance markets and to oportunistically try to scale (them) up into the supercomputer market. The “off the shelf” clusters do a very nice job for certain applications, but they don’t do a nice job for “heavy lifting” applications. And it’s particularly the “heavy lifting” kind of customers that kind of banded together in this IDC organization.
So a power shift is happening – and it’s a really healthy one, I think – where the power is shifting from large vendors who are saying, “This is what we’re going to give you, take it”, to customers who are saying, “This is what we need, make it.” I think this is the last year that we’re going to see the vendor community touting “vanilla” kinds of systems without any real differentiation.
Not everybody’s vanilla. There are vendors out there who are really trying to deliver systems for this market. As far as I know, every single vendor is supporting this – which may seem strange, but all the vendors are saying is “please be equitible. We don’t want any vendor having any particular excessive influence in this thing.”
HPCwire: Tell us about how Cray is doing since this year’s merger.
CONWAY: I think it has come along really well. I think the first months of everybody getting to know each other have gone by in an accelerated way, because there was so much that we had to do – and do together – just to make it and survive through the early months.
HPCwire: Tell us about the benefits of the new Cray SV1ex line.
CONWAY: It’s got a lot of custom engineering. It’s got processors which top out at (about) 7 gigaflops. We are not biased away from microprossesors at all. . .but really you need customize if you want systems with high bandwidth. (With) microprocessors, what improves so quickly is the transistor part of it, not the wiring part. So if you really want to have high bandwidth, you’ve got to deal with the interconnects between the processors. What Cray does is to build high bandwidth systems, low latency systems, (and) systems that are densely packaged.
Conway was later joined by James E. Rottsolk, President and CEO of Cray. Together, the two commented on Cray’s contributions to the world of HPC.
HPCwire: What do you feel differentiates the Cray architecture from other more “vanilla” architectures?
ROTTSOLK: Nearly everything. Without being too cute, we like to believe that we bring an element of innovation to the computer business. What we mean is that there’s a tendency for most people to utilize existing off the shelf technology. And it’s particularly true of most of the other computer vendors, if not all. whose focus tends not to be the high end, but tends to be desktop or midrange at best. (They) only opportunistically participate in the high end, and in doing so utilize what we would characterize as relatively “vanilla” approaches. The classic example of this is someone just taking a basic workstation “pizza box”, and rack mounting it.
We tend to look at the systems approach – to look at the customer’s problem and try to provide the customer with a new capability. And so we look at the whole system. Our T3E is a good example of this. It uses off-the-shelf microprocessors, and yet users find it to be a superior machine for a number of reasons, (including) much greater bandwidth. It has some special hardware which tends to maximize the utilization of the processors themselves, and it has sophisticated systems software. So it has utilities that other vendors don’t offer. What you find is that the utilization of this system tends to be much higher than it is with any other massively parallel system.
CONWAY: The way technologies have evolved the fastest element in the computer (in a standard sense) is the processor by far, and the trick is how does everything else keep up with the processor? And the answer is sort of in the off the shelf implementations. You’ve got great processor speeds which end up being (the) peak performance numbers.
ROTTSOLK: Take someone’s thousand processor system compared with the T3E thousand processor system, and take a look over time at how many processors where actually being used. We tend to run over 90% (usage) at major customer sites. I’ve been hearing this over and over at the show. I’ve talked to people with several large T3E installations and they’re all bragging about their 90% (or more) real utilization rates. That far exceeds whatever anyone else does.
HPCwire: Could you both comment about the proposed new performance benchmarks for supercomputers?
CONWAY: For most of the life of this industry, when people bought a multimillion dollar machine, they actually tested it. They purchased on real performance, on what it could do on their work, and the auto companies and the areospace companies – trust me on this – still do, because it’s very close to the bottom line for them. But some people have gotten away from that practice, sometimes for legitimate reasons.
ROTTSOLK: But I think the users have come to concur that the metric that has been utilized, which is this Linpack benchmark, is about as useless as to be meaningless. It bears little if any relationship to real performance.
CONWAY: It may be as many as a hundred times greater than the actual performance that we get on the system. That’s why, when you go out to the floor there (at the conference), you see not one, but every vendor now saying (they have) “the world’s most powerful supercomputer.” How can that be? Because . . .they have no standard way of measuring them. I don’t think any vendor has a choice other than to support this initiative because it’s driven by the purchasers – as it should be.
Cray computers also announced the Cray SV1ex Supercomputer product line at SC2000 yesterday. The systems are slated for availability in the first half of 2001. Publicly announced advance orders include the Arctic Region Supercomputing Center and the Department of Defense Major Shared Resource Center operated by the Naval Oceanographic Office, which plans to upgrade to a Supercluster of four coupled Cray SV1ex supercomputer systems.