Supercomputing

[HenryTo]

Cray's supercomputer at the Department of Energy's Oak Ridge National Laboratory (ORNL) upgraded (doubled) its performance to 119 teraflops - putting it number 2 on the world's Top 500 list of supercomputers:

http://www.hpcwire.com/hpc/1370386.html

[rffrydr]

With Murphy's Law in tow no doubt:

http://news.nationalgeographic.com/news/2007/04/070413-large-collider.html
_________________
Today is the Tomorrow you worried about Yesterday!

[HenryTo]

Louisiana gets 50 teraflop supercomputer - enough to put this into the top 25 in the upcoming Global Top 500 list to be published on June 27:
--------------------------------------------------------------------------------
Queen Bee

There’s a new “Queen Bee” in Baton Rouge, but this bee won’t be making speeches or signing any bills into law. What Queen Bee is expected to do, however, is vault Louisiana into the world’s upper echelons of supercomputing research capacity.

Queen Bee is of course Louisiana’s new 50-teraflop supercomputer, recently acquired to enhance the combined supercomputing capacity of Louisiana’s research universities (linked via LONI, the Louisiana Optical Network Initiative). Now that Queen Bee is up and running at the state’s Information Services Building in downtown Baton Rouge, Louisiana is expected to weigh in this year in the top 25 on the worldwide Top 500 supercomputing capacity list (www.top500.org). Queen Bee will put Louisiana in the company of such research powerhouses as the Lawrence Livermore Laboratories and Harvard University. Moreover, once Queen Bee is connected to the state’s other LONI supercomputers, Louisiana’s statewide research supercomputing capacity will stand at 100 teraflops, a massive level of processing power by any standard.

“This is a watershed moment; Louisiana has joined the ranks of the world’s supercomputing elite,” said LONI Executive Director Charles McMahon. “Queen Bee’s addition to the broader array of connected high performance computing resources through LONI marks our state’s arrival as an IT powerhouse.”

The name “Queen Bee” is a reference to the nickname Governor Kathleen Babineaux Blanco was tagged with early in her administration by a legislator who was displeased about losing a valued committee chairmanship. The Governor good-naturedly adopted the moniker as a symbol for her administration. In honor of Governor Blanco’s strong support of Louisiana’s information technology infrastructure (including a $40 million, ten-year commitment to LONI), the LONI Management Council unanimously agreed that the new “statewide” machine should be known as Queen Bee.

Benchmarking performance tests indicate that Queen Bee will operate at approximately 70% of her 50-teraflop capacity, a fairly typical efficiency level for a machine with such a high rated capacity, and, again, sufficient to warrant the expectation of making the top 25 on this year’s worldwide supercomputer rankings.

“The circuits of supercomputers like Queen Bee are commonly characterized as being ‘massively parallel,’” said Steve Landry, Vice President for Academic Affairs at the University of Louisiana – Lafayette and a member of the LONI Management Council. “The way I have explained this phenomenon in the classroom is to compare the situation to fifty men trying to move a 5000-pound slab of granite. Even if each man could individually push a 100-pound slab, when all of them try to work together to push the larger slab, there are some obvious practical issues – limited physical space for each man to get his shoulder against the stone, for example – that make it impossible to get full efficiency from all fifty men at the same time. It’s an analogous situation when parallel computer circuits are working on the same problem.”

Getting Queen Bee installed and operational in time to be considered for this year’s Top 500 list took both hard work and a little luck. At about the time the supercomputer became available to the state, Louisiana’s Information Services Building was already undergoing a electrical system upgrade to improve reliability.

“Since the upgraded control capacity was already going to be in place, [Louisiana Chief Information Officer] Rizwan Ahmed saw the opportunity to amend the facility upgrade contract to accommodate the needs of the new computer,” said Randy Walker, Information Technology Director for the Louisiana Division of Administration. “We just took a nine-month detour on the facility upgrade to get it ready for Queen Bee.”

Walker said that one of the biggest challenges in getting Queen Bee installed was providing adequate cooling, which had to be provided initially by portable chillers until the air conditioning upgrades could be completed.

Another bump in the road as the OIS team worked against the clock was negotiating around a provision in the contract with Dell (the manufacturer of Queen Bee) which tied the delivery of specific components to the degree of completion of the ISB upgrade.

“We worked until 2:00 in the morning for five or six days in a row,” Walker said, “but with teamwork and some valuable help from the Dell folks, we were able to achieve our goal. This project it great for Louisiana, and I’m proud to be able to say that I played at least small part in seeing it through to completion.”

The 2007 Top500 Supercomputer Sites rankings will be announced on June 27.

[HenryTo]

Japan aims to take back the number one spot by unveiling a 10 petaflop machine by 2009:

http://news.monstersandcritics.com/business/news/article_1316836.php/Hitachi_NEC_Fujitsu_to_jointly_develop_super-computer

To put that in perspective, the current number one on the list right now is the 280 teraflop Blue Gene system. By early next year, we should have - for the first time - a one petaflop system (Blue Gene/P). A 10 petaflop system by 2009 would really hit this out of the ballpark, so to speak, as even the Blue Gene/Q (details about this is unknown) is only supposed to perform at the 3 petaflop level. We will see.

[HenryTo]

Nvidia attempts to one-up all its rivals in the supercomputing arena with its latest product, the Tesla processor, due to be released in August. This is an amazing piece of machinery. If it is to be believed, one stand-alone Tesla GPU will have the performance of 100 high-end Intel CPUs in terms of floating-point operations per second. In "one stroke," Nvidia has upped the ante by more than seven years.

http://www.betanews.com/article/NVidia_Enters_Computer_Business_with_Deskside_Supercomputer/1182376593

Here is the press release:

http://blog.wired.com/gadgets/files/Tesla_Launch_Final.pdf

[HenryTo]

A layman's version of the article. Another must-read:

http://www.msnbc.msn.com/id/19340786/

[dash]

[HenryTo]

Dash,

Thanks for the article. I don't think a 300% to 500% increase over the next few years is too optimistic. I just got a new laptop and everything about it is fast - except for my DSL connection and maybe my hard drive. The hard drive "problem" will be solved over the next couple of years with the switch to faster solid state hard drives - so that only leaves the internet connection.

I am sending files of over 5 mb to my webmaster on a weekly basis and that is taking too long. I also have Netflix. For certain videos, you can view them on-demand - and the initial download takes over ten minutes long, which is unacceptable to me.

It also looks like Verizon is getting traction with its Fios service with over one million subscribers. Some of these folks will be utilizing as much bandwidth as they can - and the folks that deal with them on a daily basis will start demanding more bandwidth as well. With the introduction of WiMax in April 2008, this trend is inevitable.

Henry

[dash]

Hi Henry,

I misquoted him (then edited my post) - that's actually 300% to 500% per year! Bandwidth demand has been accelerating, but the latest number I heard (in March) was still 'only' about 100% per year.

There's no doubt that file sizes are getting bigger though, and bandwidth-intensive video traffic is growing rapidly, plus IPTV will probably become mainstream in the next few years. Here's a comparison of bandwidth requirements depending on file size:

email: 5 Kilobytes
Song: 5 Megabytes
Movie: 2 Gigabytes
Blu-Ray Movie: 30 Gigabytes

When the internet bubble burst, the net itself was still very much a dialup phenomenon with applications like e-mail and web browsing accounting for most of the traffic. That's no longer the case.

[HenryTo]

Dash, I guess I misread the article as well!

Difficult to see that increasing by 300% to 500% a year, unless you also make optimistic assumptions of the Chinese and Indians using bandwidth over the next few yers just like the South Koreans or the Japanese today.

Right now, total penetration in China is only slightly over 10%, and in India, only 3.5%.

http://www.internetworldstats.com/stats3.htm

Dash, do you have any reliable projections on how much internet penetration will increase in China and India over the next few years?

[dash]

Henry,

Sorry I don't have that data, but personally I think the growth in IP traffic in the US will depend more on how quickly new applications become mainstream, than on how fast internet penetration grows in China or India.

Going back to my file size example for a second:

email: 5 Kilobytes
Blu-Ray Movie: 30 Gigabytes

So a Blu-Ray movie is 6 millions times the size of an e-mail file! Traffic growth on the net is happened IMO principally because of disintermediation. So for me, the speed at which these changes happen is critical (from existing markets ---> internet based substitutes):

Consumer/Business Sales ---> E-Commerce
Physical Software Distribution ---> Network based software distribution
Music CDs and DVDs ---> Online music and video downloads
Broadcast and Cable TV ---> Streaming video

By way of example, and despite the success of iTunes, network delivery of music was only 16% of total music sales in 2006 (according to data from a Level 3 presentation). So there's a lot of room to grow!

As far as DVD distribution goes we may have reached a critical point. Total delivery cost ($ per movie) in 2007 is (from same Level 3 presentation)

Store Rental: $1.30
Mail Rental: $1.05
IP delivery: $0.20

By way of comparison IP delivery costs in 2003 were $0.95 (same as mail delivery). Clearly the more the IP delivery cost drops and the bigger the gap with physical distribution, the more compelling it becomes for the RBOCs to spend money to upgrade last mile speeds so you won't have to wait 10 minutes for your netflix movie.

In sum, I reckon that as far as bandwidth growth goes, the sooner IP video becomes mainstream, the more likely it is we see the kind of traffic numbers Chambers is claiming.

Last time I heard a stat quoted (early 2006 and by Chambers) traffic on internet backbones in Japan and S Korea was growing ~300% FWIW.

[HenryTo]

Dash,

Correct me if I am wrong, but I had thought that Mr. Chambers was talking about worldwide bandwidth growth. But then again, you are most probably correct - since bandwidth demand going forward should be driven more by U.S. apps as opposed to new internet users in China/India who are just probably going to send a few emails each day or just IM each other.

Besides streaming video, I would imagine that bandwidth would also be significantly driven by not only software distribution, but the distribution of information for business/military purposes as well - such as gene information, sophisticated 3-D designs of products as well as 3-D imaging of current products and geography, as well as your "traditional" databases. Combined with the "coming of age" of the one petaflop supercomputer over the next few years - that means that GM Chinese/Indian designers in 2012 to 2015 could design an entirely new automobile without building the physical model - using hardware and software alone - and then have the necessary bandwidth to send the design to their U.S. co-workers for review fairly quickly.

Of course, this assumption also hinges on the fact that software development will continue to improve going forward.

[dash]

This as far as I'm aware this is the world's first on demand supercomputer available for $1/cpu-hr. All you need is a credit card and a browser:

http://www.network.com/

Many non-IT companies have spent a lot of time and money building network/storage/software grids of their own, just as many factories started off by building and installing their own generators before the advent of the electricity grid during the industrial revolution. Why do it yourself when you can outsource from someone who does it better?

Also when net connectivity is faster and more ubiquitous it's got to make more sense for consumers to want to store and access all their data online (though privacy, security issues are clearly a problem). Why not put the operating system itself online? Viruses and software upgrades become less problematic, and you're computer won't crash.

I think it was Google's Schmidt who said something like: when the network becomes as fast as the processor, the computer hollows out and spreads across the network. If this is the next stage of computing then I think the issues which are important are not what operating sytem you have, or how fast is your CPU, it all comes down to bandwidth and latency.

[HenryTo]

Intel is due to release its 80-core >1 teraflop processor sometime in 2010, or even as early as 2009:

http://www.itnews.com.au/newsstory.aspx?CIaNID=54750&src=site-marq

[HenryTo]

The hits just keep on coming. Sun Microsystems announces the new Constellation supercomputing system - a machine that is capable of a peak performance of 500 teraflops. This is being deployed at the Texas Advanced Computing Center at the University of Texas and should be completed later this year:

http://news.zdnet.com/2100-9584_22-6193207.html