The supercomputer built by Virginia Polytechnic Institute last year clocked a benchmark speed at a whopping 10 trillion floating operations per second, but this year it's a no-show. The supercomputer, made up of 1,100 dual-processor Apple Computer G5 systems, was disassembled shortly after ranking as the third-fastest supercomputer in the world last fall.
Surely a supercomputer that's capable of computing 10 teraflops must have actually done something productive? Considering the short time the system was in operation, it's hard to tell whether it ran anything productive or just a benchmark test to get on a list.
I'm all for the need to demonstrate new technologies and also to understand why Congress wants the United States to have a faster computer than Japan--theirs is currently "bigger than ours." But the general emphasis on raw speed rather than productivity has a more negative effect on both the advancement of high-performance computing systems and the people who use them.
What does productivity mean when it comes to high-performance computing? It's the ratio of the number of jobs the system will run in 24 hours, or one week, or one month--or most importantly, how many jobs the system will run in its lifetime--over the total amount of resources that go into buying and running the system.
Basing decisions on this definition of productivity avoids the embarrassment of procuring "cheap" piles of hardware that never quite run your applications and end up cluttering your data center floor. Instead of procuring fast, expensive systems solely for the purpose of winning bragging rights, the focus shifts to producing meaningful computing output given limited resources--exactly where it needs to be.
Many high-performance computing systems are like the
Concorde--very fast and very exciting, but not productive.
This obsession with speed can also hold back the pace of research and scientific advancement. Supercomputing productivity determines when new pharmaceuticals are released, the accuracy of weather forecasts and even our national security. When the focus is taken off productivity, and computing speed becomes the main objective, the quality and quantity of research and development is compromised.
Compare the experience of the Concorde with that of JetBlue Airways. The Concorde, a tremendous technological achievement, was faster than any passenger-class airplane ever deployed. Whisking passengers back and forth across the Atlantic in two hours, the supersonic airplane promised to revolutionize air travel. But the Concorde was based on the assumption that faster must be better.
Then consider JetBlue, which does not tout itself as the fastest airline in the sky. Instead, the airliner focuses on productivity and efficiency by offering "red eye" flights to enable a portion of its fleet to remain productive through the night. JetBlue also schedules minimum ground time to avoid unnecessary time spent at airport gates. Quick, efficient airport turnarounds increase the number of daily flights per aircraft.
The productivity numbers speak for themselves. Last year, JetBlue served 9 million passengers and completed 67,000 flights. In its lifetime, the Concorde only served 2.5 million customers and completed fewer than 50,000 flights. The Concorde eventually failed, while JetBlue flew as an outstanding success in a largely battered industry.
Many high-performance computing systems are like the Concorde--very fast and very exciting, but not productive. Other high-performance computing systems are more similar to JetBlue, focusing on productivity and costs to yield an impressive return on investment.
For high-performance computing systems, faster isn't always better or more important. The focus needs to be on the production of systems that aren't only fast but actually get you somewhere.
Biography
Tom Quinn is vice president of East Coast operations at Linux Networx.
See more CNET content tagged:
high-performance computing,
supercomputer,
productivity,
high-performance
not a speed upgrade. VT would have preferred to use G5 XServes
at the time they built the supercomputer cluster but at that time
Apple's only options were Powermac G5s and G4 XServes. So VT
ended up using towers. Apple & VT did arrange to replace
towers with smaller XServe units when they became available.
The XServes are comparable to the Powermacs the are
replacing in performance and as far as I've heard the VT cluster
will still be composed of 1100 Macs when the upgrade is
complete. What they will end up is the same processing power in
roughly a third the area.
Either he did the investigation, but choose to leave it out because it didn't help make his case.
In this case he's not a trustworthy journalist.
Or..
He didn't do the research and didn't know.
In this case he's not a very good journalist.
With that said I see he is not an objective party at all, but a VP of a Linux firm, so I guess it's OK. Free speech allows you to trash talk your competitors all you want, while acting as if you're an objective party.
Though I do agree with the basic arguement that having the Super Computing power and not having a use for it is a waste.
This is really brought home by the metric used for the Top500 - Linpack. (See the link "mine is bigger than yours"). Linpack is a decent test, but it is far from comprehensive in measuring system throughput. Moreover, for any significant computation job, you had better know your particular usage profile. If you don't, you are likely to find that driving a Porsche in rush hour traffic is the same as driving a Yugo.
The same point is doubly true as the system perspective expands from relatively straightforward technical considerations to broader business concerns. The value of a business system depends not just on on pure speed and technical performance, but also on reliability, maintence cost, development cost, and (oddly enough) marketing.
Besides, the Russia Tu-144 was faster, and yes it did enter service. (Not for very long however.) Next time get you facts straight before you use absurd examples.
I feel better now.
Andrew J Glina
The Concorde may be an amazing aircraft; however, I wouldn't put it at in the column of greatest marvels in aviation. Afterall, the SR-71 was flying faster than both the Tu-144 and the Concorde. Tu-144 was doomed because the Soviet Union couldn't keep up with its command economy and the Concorde was doomed because it's productive days were numbered. Making the comparison the US wasn't a leader in aviation because it didn't make a supersonic transport is somewhat strange considering at the same time the US had the only space shuttle and fastest hypersonic aircraft in use. The economy just seemed to dictate there should be the Concorde and that's it. To me, I can't see where making a supersonic transport was such as great idea -- especially if the only good market available was taken.
No doubt I would give British Airways an A for effort (I wouldn't give the French any credit -- name one good primarily French made aircraft not made by another country, i.e. don't name Airbus). The Mirage? Ha. Ha. Ha. The British can make aircraft. The French are aviation freeloaders.
I do think the authors point is ok -- although I do agree with you that the comparison was apples and oranges.
The comparison of a profit making business based on proven if unspectacular technology (Jet Blue) vs glamorous technology draped in nationalist sentiment and the flag of your choice is a perfect parallel with Top500 dust collectors vs less glamorous machines which use all of their cpu cycles to produce value.
The apples and oranges disconnect is not the fault of the article, but a reflection of the Top500 promotional industry. The authors point is that we should be focusing on value (Jet Blue, Linux, and yes, even Microsoft) but we end up reading news about buzz (Concorde, Top500, .com bubble).
Job scheduling evolved in order to schedule and optimize the use of expensive computer resources toward the end of producing something. You can bet that companies that buy actual production machines based on Top500 technology will have evaluated their return on investment based on utilizing every spare CPU cycle.
Top500 bragging rights are little more than promotional offerings and welfare for engineers.
computers with which it had been built are being replaced with
G5 Xserves, and at the time the list was being compiled,
the system was down for this switch-out. Right now it's coming
back up and in testing as the hardware rotation is completing.
Check the Virginia Tech website for pictures and progress of the
new, revised system.
Because the fundamental theme of your article was based on
your incorrect original premise, your entire article is bogus.
You should be ashamed.
The underlying point of Mr. Quinn?s article is the validity of the Top500 rankings. Sure, organizations will continue to build faster and faster systems. But Mr. Quinn?s point is that why go to all the time, energy, and resources to build a 10 tflop supercomputer if its not going to be used for anything other than bragging rights? Perhaps the Top500 rankings need to be tweaked to not only show how fast but how productive the fastest computers in the world are too.
"Virginia Tech researchers are already active in a number of
areas that will benefit from the new supercomputing facilities,
says Kevin Shinpaugh, director of research and cluster
computing for the university. These include: nanoscale
electronics, quantum chemistry, computational chemistry,
aerodynamics through multidisciplinary design optimization,
molecular statics, computational acoustics, and the molecular
modeling of proteins."
"Virginia Tech researchers are already active in a number of
areas that will benefit from the new supercomputing facilities,
says Kevin Shinpaugh, director of research and cluster
computing for the university. These include: nanoscale
electronics, quantum chemistry, computational chemistry,
aerodynamics through multidisciplinary design optimization,
molecular statics, computational acoustics, and the molecular
modeling of proteins."
I call that use. This article needs to be retracted, it is not only
factually incorrect, it is borderline libellious, as there is a
suggestion that VT are spending investors money frivovously. I
am sure VT themselves would have an opinion on this.
- Missing the mark.
-
by
April 22, 2008 1:47 PM PDT
- The US Government and research institutions are interested in large, fast machines not becuase they want to run many small jobs in a shorter period, but because they have computationally intensive jobs that will not run on smaller systems. Either the memory requirement of the program is too great or the processing speed of the computer is insufficient. Supercomputers enable programs like protein folding, thermonuclear reaction simulations, large scale weather simmulations, and neural simulations. These are hard, if not impossible, to run on a collection of Intel class machines.
-
Reply to this comment
-
(19 Comments)