One of the beauties of beowulfery is that anybody can afford supercomputing when the supercomputer is made out of off-the-shelf components. Over the last several decades, supercomputers have generally been export-controlled by the fully developed countries as a weapon. This has engendered a number of ``interesting'' discussions on the beowulf list over the years.
The rationale for restricting supercomputers as a weapon has always been that supercomputers can be used, in principle, to design a nuclear device and to model explosions for different designs without the need to build an actual device. The interesting thing is that this restriction was held from the early 1970's through the end of the millenium. During that time, a ``supercomputer'' went from a speed of a few million floating point operations per second through billions to trillions. We have now reached the point where personal digital assistants have the computing capacity and speed of the original restricted supercomputers. Computers such as the half-obsolete laptop on which I am writing this execute a billion instructions per second and would have been highly restricted technology a decade ago.
Add to this mix cluster computing methodologies that combine hundreds to thousands of of GFLOPs into an aggregate power of teraflops, at a cost of perhaps $0.50 per FLOP and (rapidly) falling. Any country on the planet can now build a supercomputer out of commodity parts, for good or for ill, capable of simulating a nuclear blast or doing anything else that one might wish to restrict.
It is difficult to judge whether or not these concerns have ever had any real validity. Of course, being a physicist, I never let a little thing like difficulty stop me. In my own personal opinion, the export restrictions haven't had the slightest effect on the weapons design process in any country, nuclear or not. Nuclear bombs have never been particularly difficult to design (remember that they were originally built with early 1940's technology!). The issue has generally not been how good a bomb one can design or modeling thermonuclear blasts, but whether one can build one at all, even from a time-tested design. In a nutshell, whether or not one could lay hands on plutonium or the appropriate isotopes of uranium.
In the meantime, restricting exports of supercomputers to only those countries already in possession of nuclear bombs or capable of managing the diplomacy required to certify their use in specific companies and applications had a huge, negative impact on the technological development of countries that didn't have nuclear bombs already or the right diplomatic or corporate pull. Engineering disciplines of all sorts (not just nuclear engineering) rely on advanced computing resources - aerospace engineering, chemical engineering, computer engineering, mechanical engineering, all rely heavily on visualization, finite element analysis, simulation and other tasks in the general arena of high performance computing.
Now, was this repeated extension of the definition of a ``restricted supercomputer'' really a matter of national security and bomb design, or was it (and is it today) a way of perpetuating the control certain large industrial concerns had over the computing resources upon which their competitive advantage is based? A proper cynic might conclude that both are true to some degree, but restricting the development of nuclear bombs alone is hardly credible in a world where we have increasing evidence that any country with the will and the plutonium (such as North Korea, India, Pakistan, at the time of this writing) easily built nuclear devices as soon as they had the plutonium, and only a lack of plutonium and a war stopped Iraq.
In any event, this amusing little bit of political cynicism aside, the playing field is now considerably leveled, and is unlikely to ever again become as uneven as it has been for the last fifty years.
I have personally built a cluster supercomputer in my home that has between eight and ten Intel and AMD CPUs in the range between (currently) 400 MHz and around 2 GHz (that is, half the cluster is really semi-obsolete and none of it is bleeding edge current). Together they easily cumulate to more than a GFLOP, making it a ``restricted armament'' as of a few short years ago.
The total cost of all the systems involved is on the order of four or five thousand dollars spent over five or six years. Spending five thousand dollars all at once I could easily afford eight to ten 2.4 GHz CPUs and quite a few aggregate GFLOPS (by whatever measure you choose to use), and this is chickenfeed spent on a home beowulf.
Using even this resource in clever off-the-shelf ways, I'm confident that I could do anything from design basic nuclear devices to model/simulate those nuclear devices in action, with the biggest problem being the creation of the software required from general sources and initializing it with the right data, not any lack of power of the computers. All the components of this compute cluster are readily available in any country in the world and are impossible to restrict. Export restrictions may or may not be dead, but they are certainly moot.
At this point any country in the world can afford beowulf-style supercomputing - a bunch of cheap CPUs strung together on a network switch as good as one needs for the task or can afford. And nearly every country in the world does build beowulfs. I've helped people on and off of the beowulf list seeking to build clusters in India, Korea, Argentina, Brazil, Mexico and elsewhere. Some of these clusters were associated with Universities. Others were being built by hobbyists, or people associated with small businesses or research groups.
``Armed''18.1 with a cluster costing a few thousand dollars (and up), even a small school in a relatively poor country can afford to teach high performance computing - design, management, operation, programming - to prepare a future generation of systems managers and engineers to support their country's technological infrastructure and growth! Those trained programmers and managers, in turn, can run beowulf-style clusters in small businesses and for the first time enable them to tackle designs and concepts limited only by their imagination and the quality of their programmers and scientists, not by a lack of the raw FLOPS required for their imagination to become reality in a timely and competitive way. Compute clusters can also nucleate other infrastructure developments both public and private.
In the best Darwinian tradition, those companies that succeed in using these new resources in clever and profitable ways will fund further growth and development. Suddenly even quite small ``start up'' companies in any country in the world have at least a snowball's chance in hell of making the big time, at least if their primary obstacle in the past has been access to high performance compute resources.
In this country, I've watched beowulf-style compute clusters literally explode in ``popularity'' (measured by the number of individuals and research groups who use such a resource as a key component of their work). At Duke alone, ten years ago I was just starting to use PVM and workstation networks as a ``supercomputer'' upon which to do simulations in condensed matter physics. Today there are so many compute clusters in operation or being built that the administration is having trouble keeping track of the all and we're developing new models for cluster support at the institutional level. My own cluster resources have gone from a handful of systems, some of which belong to other people, to close to 100 CPUs, most of which I own, sharing a cluster facility in our building with four other groups all doing cluster computations of different sorts.
The same thing is happening on a global basis. The beowulf in particular and cluster computing in general are no longer in any sense rare - they are becoming the standard and are gradually driving more traditional supercomputer designs into increasingly small markets with increasingly specialized clients, characterized primary by the deep pockets necessary to own a ``real'' supercomputer. In terms of price performance, though, the beowulf model is unchallenged and likely to remain so.
Beowulfs in developing countries do encounter difficulties that we only rarely see here. Unstable electrical grids, import restrictions and duties, a lack of local infrastructure that we take for granted, theft, and graft, all make their local efforts more difficult and more expensive than they should be, but even so they remain far more affordable than the supercomputing alternatives and well within the means of most universities or businesses that might need them. As is the case here, the human cost of a supercomputing operation is very likely as large or larger than the hardware or infrastructure cost, at least if the supercomputer is a beowulf or other compute cluster.
Even with these difficulties, I expect the global explosion in COTS compute clusters to continue. Based on my personal experiences, this book (in particular, given that it is available online and free for personal use) is likely to help people all over the world get started in supercomputing, nucleating new science, new engineering, new development.
To those people, especially those in developing countries trying to overcome their own special problems with funding, with environment, with people, all I can say is welcome to beowulfery, my brothers and sisters! The journey will prove, in the end, worthwhile. Please let me know if I can help in any way.