Posted on 04/26/2002 5:55:11 PM PDT by Ernest_at_the_Beach
Caltech’s Thomas Sterling, often referred to as the “father” of Beowulf clusters, recently joined the board of directors of a relatively obscure financial firm called JJX Capital. JJX plans to build a massive AMD Athlon processor-based cluster to help it try and predict the future direction of the financial markets before its competition can. The company has labeled the machine “the largest commercial supercomputer in the world for securities trading using AMD Athlon microprocessors.” Supercomputing Online interviewed Sterling this week to learn more.
SCO: You've recently joined the board of directors of JJX Capital. Please give us a little background on the company and why you agreed to join the board.
STERLING: I was pleased to serve on the board of directors for JJX Capital because they combine long-term vision with near term practical solutions for bringing high performance computing to the important market of financial services. The JJX strategy combines my interests in Beowulf-class cluster computing (see my August Scientific American article on do-it-yourself supercomputers) with my research in Petaflops-scale computer architecture. They have synthesized a dream team of experts in the mission critical areas that promise to lead the industry in financial planning services through high performance computing.
And I'm able to bring my expertise to have meaningful impact on the success of this visionary enterprise. Joining forces with JJX Capital was a no-brainer.
SCO: JJX recently announced that it plans to build what it refers to as "the largest commercial supercomputer in the world for securities trading using AMD Athlon microprocessors." Please give us some details about this monster of a machine? Things like how many processors, interconnects, memory, peak performance estimates, a best guess on when it might be up and running, etc.
STERLING: This system will indeed be a very large Beowulf-class cluster system with in excess of 1000 nodes of dual Athlons with an aggregate memory greater than 1 Terabyte and a peak performance on the order of 2 Teraops. Assuming the business plan is followed (highly likely), they should be operational in less than six months.
SCO: To the best of your knowledge, how does JJX plan to exploit the capabilities of such a machine? How much funding (roughly) do they plan to throw at this?
STERLING: I am not liberty to release the business plan for JJX although I'm sure that Fitzpatrick would be pleased to brief you as appropriate. The goal is to provide superior financial modeling and faster response to changing conditions to deliver early financial projections; thus enabling sound decision making at the earliest possible time. This can be a powerful edge in the financial markets. I understand that JJX has a written offer for an initial funding at $25 million.
SCO: Once things are going smoothly with it, would you send me a few hot stock tips? I've been taking a beating recently.
STERLING: I'm just a wizard. If I knew of some hot stock tips, I wouldn't have to use my brains to make a living.
SCO: Are clustered systems particularly well suited to financial types of applications?
STERLING: The suitability of an application to clusters (or perhaps the other way around) depends on the degree and form of the application parallelism and the amount of communication required among the cluster processor nodes in support of the application. For clusters you want lots of coarse grain parallelism with little inter-processor communications. Financial applications in many cases exhibit a lot of coarse grain parallelism and at certain levels do not need a lot of communication. We can expect speedups between a factor of ten and a hundred from clusters (sometimes more). Beyond that, more tightly coupled parallel systems may be called upon such as symmetric multiprocessors (SMP).
Hybrid systems that are clusters of small SMPs are becoming increasingly popular so total speedups of three orders of magnitude combining both system classes may be possible. But to get that total speedup may require some significant rewriting of the application in many cases.
SCO: Petaflops-scale computing was mentioned a couple times in the JJX press release. What is your vision of a Petaflop machine? How would you build it? How long before you can build it?
STERLING: That's a tough question with a long answer. May I suggest you take a look at my Scientific American article from last July. I would either take the hybrid technology approach as described in that article or take a pure PIM approach. In either case, mechanisms for dynamic adaptive resource management are essential for latency hiding, distributed shared memory, and fault tolerance; all critical for effective computation in the trans-Petaflops regime. Five years with proper funding.
SCO: Separately, how are things at Caltech's CACR? Anything new and exciting you'd like to let the readers in on?
STERLING: As you can imagine Caltech and CACR is a pretty exciting place and a great environment in which to do pathfinding interdisciplinary research. I am currently working on a new "processor in memory" computer architecture we call a "MIND" chip for Memory-Intelligence-and-Networking-Devices we hope will provide the means of achieving Petaflops scale computing within a few years. We just held a by-invitation-only workshop, WIMPS (Workshop on the Implementation of Multi-Pim Systems), a couple of weeks ago with 50 experts at Bodega Bay to establish the opportunities and challenges of realizing Petaflops performance through future PIM architectures. Also, I've just agreed to write a 2nd edition of our recent Beowulf book and a new foundation level book on PIM architecture, both for MIT Press. So I'm keeping pretty busy, but never bored.
SCO: Is there anything else you'd like to add?
STERLING: Can't. I've got to catch a plane to the East coast. I'm giving the keynote address at an NSF meeting Friday morning.
Click here: tech_index
Who cares about the new supercomputer from Japan.
This is a Business Enterprise putting its hard earned money to work for Business reasons (greed?)
The Center for Advanced Computing Research at Caltech - CACR
I am still not sure where John and Jim are going with this new setup, but we seem to be stable at the moment.
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.