Applications
To the frontier the American intellect owes its striking characteristics.
— F. J. Turner, 1893, The Significance of the American Frontier
The feverish increase in computing power is often treated as an independent force of nature. Every 18 to 24 months, we see hardware that doubles our computation, communication, and storage, integrating to 100-fold increases every decade. We are all aware that this hardware enables new modes of operation—windows, electronic mail, digital libraries, and telemedicine are but a few examples. It also drives the scaling of physical problems, enabling new kinds of science such as the simulation of global climate modeling and the calculation of the evolution of the universe. Remarkably, the "guarantee" that waiting leads to advancement has not led to the slightest complacence in computational science. To date, the hardware speed-ups have been accompanied by comparable—if not greater—advances in algorithms.
The high-performance computer is a time machine. It brings us the 100-fold increase in computing today rather than waiting a decade. This time travel is important to many aspects of NASA's mission and defines the relationship of the Grand Challenge team to the casual simulator. The interdisciplinary Grand Challenge teams are the crews of the time machine; their captains explore science that is in the critical path today. They take the lead in defining scientific questions that cut to the heart of physical processes. They aggressively pursue new algorithms to match the 100-fold increase in computing speed to achieve a 10,000-fold increase in computational science. They define the domain of applicability of algorithms for high-quality simulation so that conventional "workstation scientists" can easily assess the quality of their own simulations. If used properly, our time machines enable the nation's research community to accelerate scientific progress, maintain world leadership, and provide long-term competitive advantages to the private sector. The knowledge gained at the leading edge becomes "conventional" within a decade.
Computers solve problems; more powerful computers solve problems of increasing complexity. Only by constantly stressing these systems to solve large complex problems do we insure sensible directions for their development. This is the crucial role played by Grand Challenge teams within the High Performance Computing and Communications Program.
In the last 50 years, a new approach to physics has arisen almost equal in important to the two "old" branches of theory and experimentation.
— Leo Kananoff, 1986, Physics Today
There are spectacular examples of advances in Earth and space science owing to numerical simulation:
The Round-2 Grand Challenge Teams are pushing these advances and bringing them to a point where they directly contribute to NASA's strategic plans. Historically, simulation enabled early exploration of physical models. Accurate simulations of physical phenomena find new roles in the critical path of NASA's science missions. Simulation frees theory from artificial simplifications and determines its nonlinear consequences after applying the projections and biases of observational (not experimental!) data. Whereas our computational power once restricted us to explore qualitative phenomena, we can now target high accuracy in a variety of scientific areas. While we will forever seek to explore qualitatively new phenomena, we can look toward teraFLOPS of computing and ask, "What simulations will solve these problems once and for all?" This stimulates new investigations of physical and numerical phenomena that can lead to deep insights. In many cases, we are forced to defer some problems and concentrate on appropriate niches where our goal can be achieved. However, with the approach to teraFLOPS and even petaFLOPS computing, the niches grow to fill the space of problems that we wish to attack.
NASA is pushing with vigor toward its reinvention as an agency that responds quickly and flexibly to new research inititiatives. Costs have become more important to insure a wide range of new missions. The critical question in reducing costs is always, "How do we pare down the scope of the mission without compromising its scientific impact?" State-of-the-art accurate simulations often provide the only answer. The life of any mission from planning to launch is long—upwards of 20 years. Even when this is reduced to the agency's goal, we will still confront the need to simulate the results of a scientific mission 5 to 10 years prior to the acquisition of data. Once again, we need our massively parallel computers to act as time machines stressed to the limit to do that which will be routine when the data is in hand.
The increase and diffusion of knowledge . . . present neither the idea of knowledge already acquired to be taught, nor of ... youth to be instructed; but of new discovery, of progress in the march of the human mind—of accession to the moral, intellectual, and physical powers of the human race—of dissemination throughout the inhabitable globe.
— John Quincy Adams, 1840, special Smithsonian Report
The HPCC/ESS Software Exchange Repository has taken a phenomenal step forward in the last year. The codes used by the Grand Challenge Teams to make their "" are available to the entire community. They are together with tools that enable one to build a supercomputer based on inexpensive commodity parts. The original goal of the massively parallel computers was to deliver fire breathing on the consumer price curve. The Beowulf Project is delivering on this promise with gigaFLOPS performance for $50/megaFLOPS, a feat recognized with the 1997 Gordon Bell award.
Versatility of the systems is shown with the new Parallel Adaptive Mesh Refinement Software Packages, PARAMESH for structured grids and Pyramid for unstructured grids. The Godiva package enables users to understand (and thereby improve) the performance of their parallel software.
Recent reports on the state of the HPCC initiative point to a need to balance "speed," measured by the gigaFLOPS brought to bear on cutting edge scientific simulations with "scale," measured by the millions of users served. This conflict is clearly seen at the National Supercomputing Centers, where they must reconcile the conflicting goals of providing the high-end resources to a small community that is pushing the envelope together with the political necessity of building a user community of tens of thousands.
When we take a hard look at the relationship between Grand Challenge teams and the wider scale use of computational science, we assert that the perceived conflict is illusory. Massively parallel computers are still hard to use. The current challenge is not just building them but programming them. It's amusing to note that vector supercomputers were once considered difficult to use and suitable only for a small range of problems. There were special courses that explained the intricacies of their programming. New graphics techniques were invented to handle the vast amounts of simulation data that they generated. Fifteen years later, nearly every workstation has a heavily pipelined architecture that has blended into the background. The software incorporates the lessons learned by a generation of scientists willing to sacrifice immediate results for longer term improvements in scientific quality that owed to the incorporation of new technology.
Our science team has demonstrated the scientific (and commercial) power of massively parallel computers. That role is exceedingly important when we examine the history of U.S. supercomputing scientific research centers, which have played an important role in insuring the preeminence of the U.S. supercomputing industry. Over the last 20 years, there have been numerous occasions when foreign manufacturers have had a lead in raw computing power. However, the large base of applications developed by U.S. scientists and engineers has insured that U.S. manufacturers remained the world-wide vendor of choice. This large base of applications led to a new generation of powerful workstations and group servers that learned from the vector architectures of the last decade's supercomputers.
Similarly, we expect that the next generation of workstations will embody many features that are characteristic of today's massively parallel computers. That is, we will see a high degree of internal parallelism in desktop machines. They will also have hierarchies of cache, RAM, and virtual memory that will use every latency hiding trick learned on today's massively parallel computers. While our Grand Challenge simulations are done on massively parallel machines, much of our analysis still occurs on workstations and group servers. By using many of the same primitives for analysis as for computing, we find that the locality enables us to run applications that used up to 10 times as much virtual as physical memory, with only the mildest degradation of performance. The generation of scientists who have become parallel programmers to accomplish their scientific goals will become a critical part of the software industry for desktop computers in the future. Each scientist participating in the HPCC program today adds to future U.S. competitiveness in the broader computer industry.
Anyone who takes a lead in a new area of computational science is beset with the responsibility of separating unique new physical phenomenal from numerical artifacts. This is an exceedingly difficult and time-consuming task. It requires a detailed understanding of algorithms and applied mathematics in addition to expertise in the physical application. Certifying algorithms, codes, and numerical results requires multidisciplinary skills. Without the careful specification of a domain of applicability, new numerical techniques and codes cannot be transferred from cutting edge teams to the large user base of "code consumers." For all of these reasons, I believe that "scale" can never be achieved without the detailed investigations that are only motivated by "speed."
Recent HPCC reports have stressed the need for greater interagency program coordination. The ESS Project has had an excellent project manager who has lead periodic meetings of all the Grand Challenge team leaders. Within the last years, the meetings have been coordinated in a way that has enabled better communication with the community at large. We have continued the tradition of having one meeting per year coincident with Supercomputing 9X. In the last year, we held a Science Team Symposium at NASA Headquarters to insure that program managers throughout the agency were aware of the progress within the Project. ESS and the CAS project held a joint scientific symposium at Ames in August 1998.
The Guest Investigator component of Round-2 has provided unique computational resources to a variety of NASA scientists. In general, results are just starting to emerge. But, we can see the phenomenal impact in one case where successful use of the Project's CRAY T3E was a critical factor in the start of NASA's Seasonal to Interannual Prediction Project (NSIPP). I believe that the work done by many of these groups will provide some impressive highlights in the next year.
The physical relations between the firmament of heaven, and the globe allotted by the Creator of all to the abode of man, are discoverable only by the organ of the eye.
— John Quincy Adams, 1840, special Smithsonian Report
NASA missions have left a considerable educational legacy. These assets are continually enhanced by the excitement of new discoveries. Grand Challenge simulations are able to animate that which appears static owing to astronomically long timescales. They can animate complex physical phenomena and illuminate the underlying simplicity of physical laws. This is becoming increasing apparent in the unique visualizations created by our Science Team.
George Lake
University of Washington
lake@hermes.astro.washington.edu
206-543-7106
http://hermes.astro.washington.edu/