|
Project COLONY Helps Brighten the Forecast for the Prediction Business
May 3, 2005 - Is it really possible to predict everything? Are there patterns in the incredibly complex, seemingly chaotic interactions between man and the elements that could one day help us predict not only the next tsunami or earthquake but also the daily weather for the year ahead, the highs and lows of the stock market for the coming decade, or the odds that a given individual would be affected by a given type of cancer over his or her lifetime.
Fifteen years ago, very few scientists held out hope that accurate prediction would ever be possible for such inherently complex events. In fact, leading mathematicians in the mid-1980s took it upon themselves to prove that some complex systems were "computationally irreducible." But now there is reason to believe that this death sentence to the prediction business may have been premature. Research in "cellular automata" is now showing that if one asks the right kinds of questions the right way, one can see broad patterns in even the most excruciatingly complex systems - enough to make it possible to predict future outcomes. One of the leading researchers in the study of cellular automata is Jim Crutchfield, who recently moved from the bastion of complex systems, the Santa Fe Institute in New Mexico, to start a new research center at the University of California, Davis. Prof. Crutchfield is a true believer, convinced that the links between past and future for virtually any system can be identified and mapped, providing a route to prediction given the right structural descriptions of emergent structures and sufficient compute horsepower to use them to perform the forecasts. That's where Sun comes in. Just as the scientific community is expecting big things from Prof. Crutchfield and his colleagues, Prof. Crutchfield is expecting big things from Sun Microsystems Laboratories (Sun Labs). In fact, he's expecting the biggest, most powerful, most complex system Sun Labs has ever produced. It's called COLONY. And it's on its way to Crutchfield's lab in the new Center for Computational Science and Engineering at the University of California, Davis. Massive Infusion of Compute Power
The study of cellular automata (CA) is highly compute-intensive. Consider a simple one-dimensional cellular automaton, for example. It has a row of cells, each of which has "states," such as black or white; and it has "rules," which define how changes in its state will influence the cell's immediate neighbors. Think of the row of cells as a miniature world; each time the rules are applied a new "year" or "generation" is spawned in the form of a new row of cells beneath the current row. You can see that even for simple one-dimensional CAs, the displayed pattern soon becomes rather complicated. You can imagine that for more sophisticated CAs involving multiple rules, such as two-dimensional or even three-dimensional CAs, the calculations become quite cumbersome. In fact, conventional computer systems simply handle them very inefficiently. According to Prof. Crutchfield, it can take a very powerful computer system many hours or even days to render a single CA simulation in three or four dimensions. The COLONY system from Sun Labs can deliver the tremendous compute power required to run CA simulations in minutes rather than hours or days. "This is the kind of power that will not only help us get our work done faster but will actually enable us to think in new and different ways," said Prof. Crutchfield. "It will help us automate the discovery of the universe." Showcase of Powerful Innovation
The COLONY system is as much of an engineering marvel as the CA simulations it will be used to process. Consider the raw numbers. The system contains up to 14 CPU boards, each of which has as many as 64 ASICs, and each ASIC has 64 CPUs, so the capacity is 57,344 processors in a single system. At peak performance, the COLONY system is capable of internally moving trillions of bytes per second between its synchronous processors. Simply put, it is one of the fastest computing systems in the world for cellular automata simulation, ranking in the top 20 fastest machines, with performance capabilities of over 14 trillion (integer) operations per second. To put that in perspective, Sun's most powerful commercial server is the 72-CPU Sun Fire E25K server, with a 150-MHz interconnect and peak performance of 115.2 GB/sec. Or to draw another comparison, the COLONY system is 10 times faster than the ASCI Blue Mountain processor, which is used at Los Alamos Labs to assess and certify the safety, security, and reliability of nuclear weapons without underground testing (the U.S. Department of Energy claimed throughput of 1.6 TeraOps for this system). While this type of direct comparison is unfair since the systems are optimized for very different purposes, the raw performance numbers give you an idea of the scale of the capabilities of the COLONY system. The COLONY system was designed at Sun Labs to provide high-performance simulation and emulation for the development, testing, and verification of large Verilog designs (the Verilog Hardware Description Language is typically used to design electronic systems). Originally, the goal was to improve Sun's time-to-market with new microprocessors by shrinking design cycle timeframes. For example, COLONY's large capacity would allow the full simulation of entire multiprocessor designs, so architects and engineers could evaluate whole systems early in the development cycle. COLONY would also allow simulations to run at much higher speeds, so problems could be detected and fixed much sooner in the design process. "This system is fully capable of running simulations of 4-way and 8-way processor designs at the same time," said Tom Riddle, lead architect on the COLONY team at Sun Labs. "It was slated to execute in only eight hours a simulation that could require two years on our 10,000-CPU Sun compute farm." In addition, the COLONY system was designed to provide its high performance at a relatively low cost. The use of both off-the-shelf components and specialized application-specific integrated circuits resulted in an innovative execution environment that delivers high performance and high reliability at a low cost compared to other lower-performance commercial options. Proving Grounds "We know the system works as expected," said Mr. Riddle. "In fact, it worked the first time we tested it, right out of the box. The system is a good fit for Jim's research in the short term. We're happy to give him an opportunity to use it to full advantage." Prof. Crutchfield was happy to oblige. "As an example of the new ideas the COLONY collaboration with Sun Labs has stimulated, I've developed a new class CA model that is adapted to describe complex molecular processes. These models would be very costly to simulate on conventional machines," he said. "The scalability of COLONY means that we can now look at CAs in many spatial dimensions without having to wait an eternity for each simulation to render. I believe COLONY will allow this new class of models to be simulated at a scale that would be relevant to biological and chemical processes." With the COLONY system currently en route to Prof. Crutchfield's lab at the University of California, Davis, it's now only a matter of time before CA simulations can be rendered in no time. And what does the arrival of the COLONY system portend for the science of prediction? "There is good cause for optimism," said Prof. Crutchfield. "I'm predicting an exciting future." Learn More: | |||||||||||||||||||||||||||