SC07: short wrap up

Next year in Austin … looking forward to it.

This year was muted relative to previous years. Some high fliers of the past were absent: I didn’t see Apple, I suspect they have decided that margins on iPhone and iTunes is simply better for them than getting in pitched battles for clusters. They haven’t had much in the way of success/installed base there. Linux Networx wasn’t there in a meaningful way (some people may have been in the whisper suites). This is due to their financial issues, which are themselves due to a combination of bad luck, and to be frank and very blunt, onerous terms and conditions imposed by federal government contracts (like it or not such T&C; do a really good job of driving small innovative companies “lucky” enough to win one of these, out of business). Winning business is one thing, winning good business is another. Linux Networx may have learned this lesson. I believe Penguin had been burned with these issues in the past at various customers in the past. If you want non-commodity like support for your commodity hardware, you had better be prepared to pay for it. Either on the purchase side, or on the delivery side. Support is expensive, like it or not, and hardware is disposable (well the commodity 1U boxen are). Until this stuff gets treated correctly, small companies may bid on government work, and be torn apart by the onerous T&C;, which are designed to inflict pain on large companies so they don’t screw over the government. Call it the law of unintended consequences. The T&C; will kill off the folks who can really help with innovative items. Appro has stepped up with its big wins recently, I hope they are not the next victims of this. The usual crowd was there: Sun, IBM, NEC, HP. Their booths were somewhat more muted than in the past. Budgets are lower. Fewer people there, though, if you watched, you saw (as usual) a high ratio of booth personnel to booth visitors. Lots of company reps were busy talking amongst themselves. The usual research/edu crowd was there, though I didn’t see some people I have known for a while. The 10GbE vendors were out, in force. A quick hyperplane is likely to be drawn between those with a clue who will be successful, and those who don’t have a clue and will look on in wonder while their customers go to those with a clue. My guess is that hyperplane is the $500/port mark. Why on earth would anyone spend $2000/port on something they can do (identically well with the same technology) for $500/port? Why would anyone not consider a technological shift to a lower cost (sub $300/port) similar/better speed technology ? 10GbE will hit big when it regularly hits and dips below $500/port, similar to IB now. Its stack is easier. Many 10GbE vendors are predicting the death of IB based upon their $2k/port pricing. Yeah. Ok. The grid. Anyone remember the hype? The next big thing? Going to supplant clusters? We should make up a hype-list, just to track “what happened to” several years down the road. The Microsoft folks were there in force. I didn’t see replacement theology. I saw interoperability. This is the message I thought they should have started with. I am happy to see this change. Generally the Microsoft people are good, and I had a nice time talking with those I saw. I suspect that some large customers told them to cut out the rhetoric, and get down to making things work together. This can and will be good. Some codes will not ever leave the OS they are running on (there is no economic incentive for the ISV to do so), and the OS needs to be part of the job description (yeah, Dragonfly can describe this, and if the system can support it … :) ) The me-too-ers were there. Pick any of the not-so-value-add resellers. They don’t engineer their own boxen, no real design. They repackage/rebrand what Supermicro and others sell them. Non-differentiated hardware/offerings. Some may not be long for this world (from private conversations, and I can’t say who …). Some folks who only engineer boxen for resale/rebranding were there, and some of these were very interesting. One in particular was an implementation of a ScaleMP system. We have a little history with ScaleMP, and I won’t go into it, but I won’t let my impressions of some of their non-technical folks color my view of the technology. This is an interesting box in that it is a great deal like the Panta boxes (Panta Systems went out of business) though not using blades. It looks interesting. Think 32 socket parallel Opteron system with 1TB ram. Yeah. 128 core Opteron single system image. I think this was in about 32U of rack space. Not sure how well it will do (SMP is hard, and SMP on a 3D grid is quite hard, though I think they are actually using a star config and IB transport). Higher latencies than SGI’s, but darn it, it looks neat. Evergrid (curiously who hired Fred van der Bosch from Panta after Panta folded) was there. I saw their stuff. It is good. It is cool. Everyone should want it on their clusters. Everyone. They provide one of the major missing/critical pieces. Yeah, it is that good. No we are not resellers, owners, etc. Yes, we are looking at how to do this (reseller/integrator). I sat in on the Cell BOF with Vipin and a quite a few others. Dongarra showed off iterative refinement (use single precision for bulk O(n^3) calcs, and use double precision to clean up the results for O(n^2) correctors). Works in some cases, but you need some sort of OOM difference between algorithms for this to be meaningful. He showed it in the context of a Cholesky decomposition and a QR and LU factorization. This was a good conference. Less hype, more results. BTW: Accelerators were everywhere. SGI had our partner Progeniq in their booth for rendering. Others were showing off accelerator boxes. DRC, extremedata, and others were talking about their bits. Intel talking about FSB units. IBM talking about Roadrunner (large scale accelerated supercomputer) And so on. When RoadRunner finally boots and runs, it likely will be the number 1 spot on the top 500 for a while. And it will get there with accelerators. 32k Cell chips. A theme running through discussions with people has been that some of the IDC numbers for some markets may be overstated. The purchases we and multiple other folks see don’t seem to match the IDC numbers (percentage wise). Our discussions with people in this market suggest they have great needs and very little money. Which flies in the face of what IDC reports. I suspect that there may be a little … rebranding … of what a system is. Yeah, lets call this big Oracle box a supercomputer, so we can say we do HPC. I think IDC needs to delve into this in a little more depth, as I think there are multiple signals in this data. I might bug a contact I have there about this, and ask them to look into it. AMD. Worth a short take. Spoke with quite a few AMDers. I have a sense that they are frustrated about not being able to deliver Barcelona in meaningful volumes. I noticed that Dell has effectively taken off “quad core supported” from its SC1435 unit when ordering. AMD needs, badly, to clean up its delivery. And the folks there know that. And they can’t do anything about it. It is at this point in time, I usually point out something that will seem unrelated. But it isn’t. The AMD CEO needs to read a little story Arthur C Clarke once wrote. Clarke claimed it was required reading for MIT freshmen for some period of time. It should be required reading for AMD management. It is all about how to lose a war by having superior technology. Its a 20 page short story entitled Superiority. AMD would do well to heed the message within it. I commented to one AMD person that AMD needs to be first out with 8 core. And curiously, it would be real easy for them to do it. They have all the technology in place. Two dies connected via HT links, and one set of pins. In an MCM carrier. Penryn and its followons will get there. AMD does not want to let Intel get there first. I saw the exludus booth, but no Stephan or Benoit, so I didn’t get to say hi. I looked all over for the Ibrix booth and couldn’t find it. All in all, this was a good conference.