Slashdot Mirror
DARPA Awards HPC Contracts To IBM, Cray, Not Sun
snedecor writes "DARPA has awarded a third round of funding for the next-generation petascale computing system. IBM and Cray roughly split the $494M, while Sun, with little track record, received none. This is in spite of Sun's radical proposal for proximity communication."
You can pet a dog, you can pet a cat, but you can't petascale.
You can't talk about Wikipedia's flaws on Wikipedia
IBM's radical idea was to release whatever product they'll have ready in 2010, but with enough processors to reach petascale. Go Blue! Had Sun offered up Niagara2 @ petascale, then they might have had a chance to win as well.
Now, of the two stated applications, which do you think is more interesting to the military? I suppose one could argue defense against bioterror, but it's still kinda scary.
Google's HPC "innovations" are surely software-based, are they not? Not sure they've done anything in the hardware arena that warrants press releases.
The HPCA program is a "cover" (not so cover) funding to companies. The problem is that it is not so clear that it is even good for them. The reason is that "lots" of additional resources from these companies are also diverted for these projects. Since these machines have a "doubtful" application besides the DARPA contract, I think that it may be better for these companies to invest on research more related to their product or may-be products.
For example, Sun Labs was in charge of the DARPA project at Sun. They have "invested" 3 years on that. My question is "what do they have to show?".
They do not have publications on any top computer architecture conference, they do not seem to have anything that may save Sun ass. (At least from
an architecture point of view)
This is not such a strange comment, I have head it from people at IBM research itself. Some people there is not sure that winning is the best thing either.
Google are experts at scaling out, not up. Running applications to serve 10000 users at the same time is different from executing one massive program. The main way in which Google and HPC (High Performance Computing, i.e. Supercomputers) overlap is in enormous file systems.
Niagra and Niagra 2 have lousy floating point performance (1 FPU for entire chip, shared by all the cores). Given that the DARPA project is for FLOPS, Nigra is just about the worst processor one could propose for the project.
I love the Niagra design; for 90% of what I need done, it does great. It's just terrible at floating point.
Sometime down the line, past Niagra 2, one could posit a version of such a chip with enough floating point units that it's efficient in FLOPS; it's an obvious upgrade of the current chips. However, that also is not optimal for FLOPS in the HPC regime. HPC is all about hiring enough computer scientists and physicists to micro-optimize the code so that you make close to theoretical maximum efficiency in utilizing the CPU cores. Niagra is all about keeping enough contexts on the chip that you can productively use the time that normal programs spend wasted, waiting for main memory accesses and so on. HPC by definition spends the CS time and effort to avoid that already.
Google's advances are very far from traditional HPC applications like fluid dynamics, weather forcasting, solid body simulations, waveguides, thermal reactions, particle dispersion, oil discovery, etc. Google does data mining, and transactional processing. The very problem that the darpa HPCS program addresses, is that the bulk of the HPC systems sold in the US are just clusters of off-the-sheld database/web-optimized servers. It turns out that these clusters don't deliver very high levels of efficiency, either computationally, or from a power/cooling perspective. Google rolls their own servers, but they still fit into the database/web-optimized server camp. Their software acheivements are all in the data-mining category.
This is not to say that the defense department doesn't need lots of high-end database servers. They use them by the truckload. However, the need for advances in this area are being met by the hardware and software markets. Market forces were not, however, stimulating truly interesting research at the high end of the HPC marketplace. Thus the DoD needed to put together this competition.
Someone else will probably license Sun's proximity communication technology at some point, but it might be the graphics card makers. The proximity communication stuff lets you hook together multiple chips, almost as if they were part of the same die, using a bunch of capacitor-like plates in the chips. This could be very useful for putting some amount of memory (almost) on the cpu die, and putting a very wide bus between that memory and the processor. Both IBM and Cray currently use very expensive ceramic multichip modules to connect multiple dies together, and they are still somewhat limited in the number of connections that can be attached through the modules.
Apart from that, I don't really know what advances they had. Solaris can scale to 100 processors fairly well, but both IBM and Cray have been working on scalable operating systems for systems with tens of thousands of CPUs. The rock processor would likely be a lot faster than sun's current processors, but it's an incremental advance for microprocessors. Both IBM and Cray are working on more radical technology with FPGAs, vector CPUs, highly treaded designs, and sophisticated coprocessors, and very scalable interconnects.
Niagra 1's FP indeed sucks, but I thought they were fixing that for Niagra 2. I heard about an FPU-per-core design being done (maybe that's a later follow-on?).
H-P and SGI were knocked-out at the end of round one; Cray, IBM & Sun were selected
for round two.
Once you start searching for US chip design and manufacturing firms, you realize that there are tons of them that produce silicon that is general purpose. You only listed the three biggest.
It is your personal duty to fight for what is right on a daily basis. Ignoring injustice is identical to approving
HP couldn't figure out how to sell printer ink for a supercomputer. Thus, they have no way of making any money on it.
This was modded funny, but it's an interesting point. I've been able to tour a couple big HPC type facilities and I've spoken to some academic types who were involved in other projects. From what I can tell, the HPC community is all over Linux. Linux is pretty much the de facto platform of choice, for all sorts of reasons. I understand that a lot of people in the life sciences were enamored of Apple's later-model PowerPC hardware for HPC applications, but even they would tend to reformat and install Yellow Dog. Wouldn't it be a shame if Sun got overlooked for this contract because it insisted on pushing Solaris?
... because OpenSolaris is nearly as free (in every sense) these days as Linux, and it has some great technology in it, including a battle-hardened networking stack.
... too bad it came to naught.
And then again, might it not be a shame again
A poster above questioned whether the companies that get awarded these DARPA contracts really see any benefit from it. But a major government HPC contract running on Solaris could potentially be a big marketing boost for Sun
Breakfast served all day!
It's Niagara
Niagara II (T2) has one floating point unit per core...so for a T2 outfit with all cores, eight FPUs.
Fortress, the language being developed by a bunch of people led by Guy Steele, was funded as part of the HPCS effort. This means that DARPA is going with IBM or Cray's language (X10 for IBM, Chapel from Cray). According to a press release quoted at http://www.eweek.com/article2/0,1895,2063043,00.as p (but not available at http://www.sun.com/smi/Press/sunflash/index.xml) the work will continue, but how likely is it to succeed?
l Slides9Jun2006.pdf - I thought it was pretty impressive.
Guy Steele gave an excellent talk at OOPSLA on Fortress - the slides are at http://research.sun.com/projects/plrg/PLDITutoria
The groups's site is at http://research.sun.com/projects/plrg/
http://www.acooke.org
Sun's HPC contribution is in optical chip interconnects, described (somewaht fluffily) at http://research.sun.com/spotlight/2006/2006-04-07_ Sun_on_HPCS.html
davecb@spamcop.net