Fifty Years of Computing at LLNL as a lens to the futureHeidelberg 22 jun 2002 Looking back to 50 year Lawrence Livermore National Laboratory, Dona L. Crawford, Associate Director of Computation, gave an outlook into the year 2005. She listed important issues for a scientific computing centre.
"True, we may not know the future in detail, but certain issues existed already 50 years ago. Those issues persist today and will accompany us into the future", explained Dona Crawford. Different components influence the LLNL Computing and Computational Science. It consists of Culture and Organisation, Balanced Funding, Science and Engineering, Verification and Validation, Software Infrastructure with the elements Modelling Methodologies, Algorithms and Programming Environments as well as the Hardware Infrastructure with Platforms and Shared Resources. Balance is crucial within and across these dimensions. She then discussed these topics in detail.
Simulations play a key role in virtually every LLNL Programme. Experiments are impractical in the field of Environmental with global climate and groundwater flow and in Physics & Biology with materials modelling and drug design. Experiments are prohibited in Stockpile Stewardship with radiation transport and hydrodynamics. Experiments are too expensive in Lasers & Energy with combustion and ICF modelling and in Engineering with structural dynamics and electromagnetics.
In 1952, LLNL was founded by Edward Teller and others whose experience and interaction with such researchers as John von Neumann reinforced the importance of coupling theory to experimental computational capabilities. Lessons learned early and reinforced over time create today's culture and organisation and that is needed for the future. There is a belief in the value of computational modelling. Hire and retain those with computational interests and commitment, which means that teams are key and multi-disciplininary teams essential.
On the hardware side, vendor partnerships help accelerate capabilities. In the past there was R&D funding to develop prototype systems (e.g. IBM Stretch, IBM Photostore). LLNL is committed to buy first or early production systems (e.g. CDC 6600, 7600, Star, Cray 1, XMP, YMP, Cray 2, BBN Butterfly, Meiko). Actually the present systems follow the ASCI Pathforward strategy - interconnects, rendering, storage. They have the largest machines such as ASCI Blue-Pacific, ASCI White and in the future will support IBM Blue Gene-L and ASCI Purple and beyond. The slide with the supercomputing history of LLNL is summary of all the supercomputers of its time, starting with a Univac 1 in April 1953, IBM Stretch in March 1961, CDC 7600s (1st of 5) in March 1969, CDC Star-100s (1st of 2) July 1976, Cray 1s (1st of 3) September 1978, Cray 2s (1st of 3) September 1985 and so on. ASCI purple and BlueGene/L are expected in 2004.
On the algorithmic side - computational mathematics, improvements are 10 to 100 fold. She presented an example of the importance of optimal algorithms. It is sometimes argued that increasing power from architecture diminishes the importance of algorithmic research, but the opposite holds. The more powerful the computer, the greater the importance of optimality.
She supposed Alg1 solves a problem in time CN2, where N is the input size, and Alg2 solves the same problem in time CN. Suppose that the machine on which Alg1 and Alg2 run has 10,000 processors, that have been parallelised to run in constant time (compared to serial time), Alg1 can run a problem 100X larger, whereas Alg2 can run a problem 10,000X larger, in 2D 10,000X disappears fast, in 3D even faster. Alternatively, filling the machine's memory, Alg1 requires 100X time, whereas Alg2 runs in constant time. Large 10,000- processor machines are expensive, and optimal algorithms are the only algorithms that we can afford to run on them.
Dona L. Crawford gave an example of potential run time improvements. An application impact example: the recent major model run on 1.5TF capability took about 40hrs, on a 10 exp(23) system, it would take ~2µs. She extrapolated from 1990 and gave an outlook on 2050.
For every 1 gigaflop peak performance, we need:
Where 10 exp(6) = mega, 10exp(9) = giga, 10exp(12) = tera, 10exp(15) = peta, 10exp(18) = exa, 10exp(21) = zetta, and 10exp(24) = yotta. Thus we learned new words for unbelievable dimensions. Exaflops only scratch the surface of potential for biological modelling:
Dona L. Crawford summarised that experience at LLNL underscores that the key issues important 50 years ago are still important today, and will be important into the future, only the details will change. LLNL keeps focusing on pushing the frontiers of HPC balanced hardware and software infrastructure. There are strong personal and organisational commitment to the computational science research programme. They will push hard on what can be modelled, and on model accuracy and the scientific and modelling culture and organisation - teams and collaborations. They will improve verification and validation methodology and need an adequate, balanced funding. It is important to achieve balance within and across all the dimensions discussed in the beginning of this article.
Uwe Harms |
|||||||||||||||||||||||||||||||||||
 |