Journal of the Earth Simulator, Volume 2, January 20051Science-Driven System Architecture: A New Processfor Leadership Class ComputingHorst Simon1, 2, 3*, William Kramer2, William Saphir2, John Shalf3, David Bailey2, Leonid Oliker3,Michael Banda1, C. William McCurdy4, John Hules1, Andrew Canning3, Marc Day3,Philip Colella3, David Serafini3, Michael Wehner3and Peter Nugent31Computing Sciences Directorate2National Energy Research Scientific Computing Center (NERSC) Division3Computational Research Division4Chemical Sciences Division(Received November 4, 2004; Revised manuscript accepted January 6, 2005)Abstract Over the past several years, computational scientists have observed a frustrating trend of stagnatingapplication performance despite dramatic increases in peak performance of high performance computers. In2002, researchers at Lawrence Berkeley National Laboratory, Argonne National Laboratory, and IBM pro-posed a new process to reverse this situation [1]. This strategy is based on new types of development partner-ships with computer vendors based on the concept of science-driven computer system design. This strategywill engage applications scientists well before an architecture is available for commercialization. The processis already producing results, and has further potential for dramatically improving system efficiency. Thispaper documents the progress to date and the potential for future benefits. An example of this process is dis-cussed, using IBM Power architecture with a computer architecture design that can lead to a sustained per-formance of 50 to 100 Tflop/s on a broad spectrum of applications in 2006 for a reasonable cost. This partner-ship will establish a collaborative approach to modifying computer architecture to enable heretofore unreal-ized achievements in computer capability-limited fields such as nanoscience, combustion modeling, fusion,climate modeling, and astrophysics. Keywords: high-performance computer architecture, science-driven computer design, Blue Planet, VirtualVector Architecture (ViVA), application accelerator1. STRATEGIC APPROACH TO A LEAD-ERSHIP COMPUTING TECHNOLOGYThis paper presents a plan that will maximize thereturn on the U.S. government’s investment in high per-formance computing, initiate a new wave of scientificdiscovery, and enable the solution of problems of nationaland global importance. Our vision is guided by the fol-lowing analysis:1. Government investments, such as the U.S.Department of Energy’s (DOE) Leadership ClassComputing project, must lead to widely deployablenew technology for high-end scientific computing. Ifsuch an investment leads merely to a series of exper-iments or the purchase of a single machine, it willnot have a lasting impact. 2. The technology needed will not spontaneouslyappear on the market. By taking a passive approachthat relies on evaluating and procuring existing ven-dor offerings, the high performance computing com-munity has ceded leadership to other requirementsthat are increasingly incompatible with the needs ofhigh-end computing. 3. Several national panels have concluded that the rulesof engagement between the scientific communityand the American computer industry must be revised[2, 3, 4]. Scientific applications must directly influ-ence machine design in a repeating cycle: (a) scien-tific applications input, (b) computer design withincreased performance, (c) deployment and deliveryto the scientific community, (d) repeat.4. Successfully changing the rules of engagementrequires partnerships with the American computercompanies with the resources and the track recordsof research and development in high performance* Corresponding author: Dr. Horst Simon, Lawrence Berkeley National Laboratory, One Cyclotron Road, MS 50B4230, Berkeley,California 94720, U.S.A. E-mail: [email protected]. Simon et al.2J. Earth Sim., Vol. 2, January 2005computing. To justify the necessary commitments,there must be a national consortium of laboratories,computing facilities, universities and researchersequally committed to changing the future of thecomputing capability available to the scientific com-munity. 5. Evaluating a representative array of applications toestablish precisely their algorithmic characteristicsprovides a clear understanding of the limitations ofcurrent high-end systems of all designs, from clus-ters to vector computers.6. Over the past two years, the Blue Planet partnershipled by Lawrence Berkeley National Laboratory(Berkeley Lab) has worked closely with IBM todesign a machine that better meets the needs of sci-entific applications. The goals and methodology ofthis partnership were validated by the successfuldesign and implementation of the $100+ millionASC Purple machine at Lawrence LivermoreNational Laboratory, based on the Blue Planet nodedesign. It is the first success of this science-drivendesign process.2. SCIENTIFIC APPLICATIONS ANDUNDERLYING ALGORITHMS DRIVEARCHITECTURAL DESIGNThe central goal of this strategy is to deliver new scien-tific results on computations of a scale that greatlyexceeds what is possible on current systems. It is possi-ble, within a reasonable cost, to create by 2006 a systemwith sustained performance rates of 50 to 100 Tflop/s onscientific applications of national and global importancefor an acceptable cost. We have identified the followingexample application classes as being ripe for break-through science using very high-end computing, and rele-vant to some of the most important national objectives:nanoscience, combustion modeling, fusion energy simu-lations, climate modeling, and astrophysics. Table 1 sum-marizes the goals, computational methods, and exampleapplications of each science area.The most effective approach to designing a computerarchitecture that can meet these scientific needs is to ana-lyze the underlying algorithms of these applications, andthen, working in partnership with vendors, design a sys-tem targeted to these algorithms.Table 1 Science breakthroughs enabled by leadership computing capability.Science Areas Goals Computational Methods Examples of Breakthrough ApplicationsNanoscienceCombustion ModelingFusion EnergyClimate ModelingAstrophysicsSimulate the synthesis and predict the properties of multi-component nanosystemsPredict combustion processes to provide efficient, clean and sustainable energy Understand high-energy density plasmas and develop an integrated simulation of a fusion reactorAccurately detect and attribute climate change, predict
View Full Document
Unlocking...