Supercomputing and China
written_testimonies
before the US.-China
Economic and Security Review Commission
Field Hearing on “China as an Emerging Regional and Technology Power: Implications for U.S. Economic and Security Interests”
February
12-13, 2004, San Diego, California
Dr. Francine Berman
Director, San Diego Supercomputer Center
Professor and High Performance Computing Endowed Chair, U.C. San Diego
Mr. Chairman and Distinguished Commissioners:
I am pleased to testify today before the Commission on developments in supercomputing and science. Understanding today’s integrated and global approach to technology and science has important implications to U.S. technology policy towards China.
I will begin by discussing the area of supercomputing and the role it plays in today’s science and technology landscape to provide context for these remarks. I will then discuss supercomputing in China. I will complete my written_testimonies with some remarks on leadership in science and technology today.
Supercomputing Today
“Supercomputing” is the highest performing technology that can be brought to bear on an individual problem. Over the last 20 years, supercomputing technologies have been widely acknowledged as an enabler for new generations of scientific discoveries. The increasing capability for large-scale analysis, modeling, simulation, and other key approaches have enabled dramatic progress in a wide spectrum of scientific disciplines and fundamental discoveries in science and engineering.
Today, supercomputing is accomplished by more than high-performance computers. Modern “supercomputing” applications utilize high-performance computers, high-speed networks, large-scale data storage, scientific instruments, and other resources, often in a coordinated way, to achieve breakthrough scientific results. For example, the recent sequencing of the Human Genome combined special-purpose machines for high-throughput DNA sequencing with large-scale data storage and high-performance computers to calculate 500 million trillion base-to-base comparisons. Over 80 trillion bytes of data were produced by the DNA sequencing pipeline and were utilized in a genome analysis calculation. The calculation required over 20,000 CPU (central processor unit) hours, arguably the largest computational biology calculation ever performed at the time [1].
Coordination of resources for the execution of an individual application is the focus of “Grid Computing,” a software and services technology which allows the virtualization of distributed resources. The potential of Grid technologies to link distributed resources for computation, data management, and multi-resource supercomputing has resulted in a global effort in the development and deployment of Grid software and services over the last decade. Projects such as the U.S. TeraGrid [2], the Pacific Rim PRAGMA Grid (of which China is a founding member) [3], and the European Union’s DEISA (Distributed European Infrastructure for Supercomputing Applications) project [4]demonstrate the potential of Grid Computing as a unifying global concept.
Most commonly, Grid technologies are used to link stand-alone, high-performance supercomputers and data storage, scientific instruments, visualization facilities, and other resources. A few applications can achieve supercomputer-level performance using collections of less powerful resources alone, for example, SETI@home achieves tens of TeraFlops on millions of laptops and PCs [5,6]. However, most supercomputing-class applications are not able to tolerate the longer latencies, heterogeneity, and dynamism of very large collections of small-scale resources. Instead, they use the Grid to link a few high-performance resources, coupling the fast interconnects and optimized architectures of today’s high-performance supercomputers with remote storage or data-generating scientific instruments across high-speed networks to enable breakthrough results that were not possible a generation ago.
Even in the most traditional conceptualization of supercomputing, high-performance supercomputers themselves are aggregations of key technologies from a global spectrum of sources. Today, Chinese supercomputers use imported Itanium chips from Intel and Opteron chips from AMD. U.S. supercomputers are made by multi-national corporations such as IBM, Cray, and Hewlett Packard, and incorporate RAM made in Korea and other semiconductors from Malaysia, Singapore, Taiwan, Japan and elsewhere, as well as components made in the U.S. Over the last few decades, a number of different architectural models for high-performance supercomputers have been developed, reflecting complex trade-offs between differing technologies and various alternatives with respect to the amount of supercomputer “real estate” devoted to processor power, memory and interconnect.
Leadership in high-performance supercomputers today is typically marked by one’s rank in the global Top500 list (www.top500.org) [7,8]. Of the world’s “fastest” 15 machines (measured in terms of speed when executing the solution of a dense system of linear equations on a dedicated machine), China’s submission is number 14, preceded by thirteen U.S. and Japanese supercomputers, with the top spot occupied by Japan’s Earth Simulator. The machines at the very top of the Top500 list are generally interconnected aggregations of “smaller” machines, differing in CPU instruction set and high-speed interconnect. Most are programmed using message passing, making them similar for the programmer. Vector supercomputers have higher memory bandwidth than cluster supercomputers, and address a wide class of applications, however vector supercomputers are generally significantly more expensive than cluster supercomputers. Probably the largest commonality between the machines at the top of the Top500 list is the considerable commitment, planning and funding over a substantial period required to develop and support the human, software and hardware infrastructure required for a leadership position.
In brief, although stand-alone high-performance supercomputers remain at the core of modern supercomputing, today’s applications use a broad spectrum of high-end resources, integrated by software technologies, and coupled to support new generations of advances and discoveries.
Science without Boundaries
Concurrent with the globalization of technology, science today is conducted as a “team sport.” For the last two decades, the most challenging science and engineering problems have been tackled by (often international) teams of researchers drawn together by common interests and expertise. Competition between groups and schools of thought is typically much more relevant to scientific discourse than competition based on national associations. For example, the 2000 Gordon Bell Prize, established to reward practical uses of parallel processing and given for the best performance improvement in an application, was won by an international team including researchers from the Max Planck Institute for Gravitational Physics in Germany, and researchers in the US. The prize was awarded to the team for using a globally distributed collection of supercomputers to solve Grand Challenge problems in Physics [9]. The Gordon Bell competition is part of the first tier “Supercomputing” SC Conference, a technology and computational science-oriented meeting that has become a truly international venue. SC integrates globally linked activities during a week-long technical meeting and exhibition, guided by an international technical program committee which in 2003 included distinguished scientists from 10 different countries [10].
Apart from individual efforts, many scientific disciplines are coming together as global communities in unprecedented ways to use technology to enable new discoveries. In High Energy Physics, scientists from around the world are coming together through the Compact Muon Solenoid (CMS) project, an international effort to upgrade the Large Hadron Collider (LHC) at the CERN particle physics laboratory in Geneva. The CMS project is one of the largest international scientific collaborations in history. Its goal is to bring protons into head-on collision at higher energies (14 TeV) than ever before achieved. The CMS experiments will allow scientists to penetrate further into the structure of matter and recreate the conditions prevalent in the Universe just 10-12 seconds after the "Big Bang". As of mid-2003 2,300 people from 159 institutes in 36 countries, spanning Europe, the Americas, Asia, and Australia, were involved with CMS [11]. In September 2003, the project launched the LHC Computing Grid (LCG), a network of computers and instruments designed to handle the unprecedented quantities of data that will be produced by experiments at the facility. More than 12 petabytes of data -- 12 million gigabytes -- will be generated each year, the equivalent of more than 20 million CDs. The LCG will meet the computational challenge of analyzing and mining this data by deploying a worldwide computational Grid, integrating the resources of scientific computing centers spread across Europe, America and Asia [12]. Analysis of this data will enable breakthrough discoveries critical to the entire High Energy Physics community.
Similarly, the Astronomy community is using the International Virtual Observatory Alliance to federate sky surveys from large-scale telescopes all over the world, including telescopes in the United States, Chile, the Canary Islands and Australia. The effort will provide a comprehensive whole-sky data collection that can be mined and analyzed for new information [13]. Federating key astronomical data will make it possible to undertake studies that would otherwise require so much time and resources that they would effectively be infeasible. The ability to correlate massive data sets over a broad range of wavelengths, spatial scales, and time intervals will make it possible to efficiently search for rare and/or complex types of astrophysical objects and phenomena, and have already produced new discoveries.
Today’s team-oriented, global and virtual approach to science and technology provides tremendous potential for advances and discoveries, but is not well-represented by traditional notions of ownership, credit, etc. It is within this landscape, that China’s evolving science and technology programs must be considered.
Supercomputing in China
China’s rate of growth in supercomputing has been rapid. China has several supercomputer centers, the largest of which is CNIS (Computer Network Information Center) at the Chinese Academy of Sciences [14,15] which currently has a staff of roughly 170. By comparison, NSF’s supercomputer centers in San Diego (SDSC) [16] and Champaign-Urbana (NCSA) [17] have roughly 400 staff and PSC in Pittsburgh [18] has roughly 90 staff.
China has the domestic capability to build fast supercomputers and its current most powerful supercomputer is based on a commodity cluster model. The Chinese have nine supercomputers on the current Top500 list with the most powerful supercomputer, the DeepComp 6800, ranked 14th as of November, 2003. DeepComp 6800 is built by the Chinese Legend Group Corporation [19]. The remaining 8 supercomputers are numbers 82, 90, 141, 163, 188, 435, 443, and 455 on the list and are built by Legend, IBM, HP and at Shenzhen University. China’s first appearance on the Top500 list was in 1995 with an IBM computer at the China Meteorological Administration ranked 143. It was not until 2002 that the Chinese had a computer ranked higher (number 43 at the Chinese Academy of Mathematics and System Science), and on the most recent list China’s submission occupies the 14th spot, preceded by machines deployed in the U.S. and Japan.
In addition to high-performance supercomputers, China is also moving into the area of Grid Computing. With the U.S. and others, China is a founding member of Pacific Rim Applications and Grid Middleware Assembly (PRAGMA) [20], and is developing its own national Grid projects, including the China National Grid (CNGrid) consortium [21] and the China Grid Forum [22]. China also participates in a number of collaborative network projects including the GLORIAD project [23], which is linking U.S., Russia, and China in a global-ring network used for joint scientific and educational projects, the Asia-Pacific Advanced Network (APAN) [24], Asia-Pacific Grid (ApGrid) [25], and others.
Technology is receiving a big push in China, as it is all over the world, by access to and availability of information through the Internet. China’s increasing interest in technology can be seen by its jump in Internet usage: In 2003, China’s Internet usage rose 34.5% to 79.5 million web users, exceeding Japan for the first time, to claim the most Internet users in Asia. During that same year, sales of computers in China reached 17 million in the first nine months, nearly twice the 9.17 million sold in 2002, according to industry statistics [26].
The rapid march of technology is perhaps the largest factor affecting supercomputing in China. With the success of commodity cluster supercomputers, which China can build, and the growth of Grid Computing, China has a growing supercomputing capability, increasingly independent of US export policies. In the areas of supercomputing and science, intellectual property protection and limited market access will not necessarily deter China’s progress. The growth of open source software, and the scientific traditions of open exchange of information and international collaboration are the principal factors enabling science and technology in China, and as they are all over the world.
What does it take to be a leader in supercomputing?
More than any other factors, the key to leadership in supercomputing is the ability to make it a national priority, and to back it up with resources, commitment, and planning. For traditional high-performance supercomputers, the top spots on the Top500 list demonstrate a multiplicity of architectural models, uses, and institutional environments. What is common among all of them is the commitment of resources applied to their successful development and deployment.
Consider the Japanese Earth Simulator, number one on the Top500 list since June, 2002. Project planning started in 1997 and was funded at a U.S. equivalent of almost $500 million for hardware development and deployment, housing and maintenance, human support and development, and scientific research. The Earth Simulator architecture is based on a parallel architecture model that has proved successful for high-performance codes for several decades. NEC’s version for the Earth Simulator incorporates custom-designed chips capable of sustained performance of 35 trillion floating point operations per second (as measured using the benchmarks of the Top500 list). The scientific models run on the Earth Simulator are optimized community codes used by the ocean and atmospheric global modeling community as well as the solid earth community. The project required an investment of hundreds of millions of dollars and sophisticated planning, organizational, and support infrastructure for a sustained period of years [27].
Today, many countries are working towards global leadership in science and technology and are building large-scale national efforts to achieve that leadership. For example,
· The European Union has embarked on large-scale pan-European projects to build high-performance and Grid technologies. Two major Grid efforts funded last year include DEISA (Distributed European Infrastructure for Supercomputing Applications) [28], which is connecting existing supercomputing centers into a Grid, and EGEE (Enabling Grids for E-Science in Europe) [29], which involves 70 institutions in 27 countries and is focusing on Grid middleware. These two large-scale efforts, with total funding of approximately $63 million, include EGEE funding of $38 million over two years for software development alone, involving “human infrastructure” of some 200 FTE (full-time equivalents).
· Japan is following its success with the Earth Simulator with a plan to build a very large 25+ teraflops Grid cluster connected by high-bandwidth networks. Vendors include IBM, which will provide the world’s most powerful Linux cluster supercomputer, to be integrated with other systems to form a massive, distributed computing Grid. The Japanese Grid is intended to enable collaboration between corporations, academia, and government and to support research in a variety of areas including Grid technologies, life sciences bioinformatics, and nanotechnology. [30]
· Three years ago, the UK initiated a concerted effort to develop leadership in Grid Computing and middleware. Today, the UK e-Science program, funded at more than $220 million over 3 years, is providing global leadership in many areas, including operational Grids and a major effort in open middleware (OMII) involving industry and the academic and research communities. The OMII vision, funded for 3 years at almost $12 million, is to become a source for reliable, interoperable, open-source middleware, and is focused on moving research quality software into reliable, robust, usable software [31, 32].
Technology leadership in supercomputing rests on the ability to successfully fund, deploy and integrate resources and infrastructure at the largest scale; science leadership rests on the ability to do breakthrough work. Today, science and technology leadership have become intertwined, and the advances of each bootstrap the other to new heights. Both the United States’ and China’s success will be incumbent on an ability to bring national strengths and serious resources to build leadership. In today’s increasingly virtualized and integrated world, everyone is both a potential resource, and a potential competitor.
Thank you for the opportunity to testify today.
References
1. Celera Genomics Completes the First Assembly of the Human Genome, http://www.celera.com/celera/pr_1056581295
2. TeraGrid.org, http://www.teragrid.org/
3. Pacific Rim Applications and Grid Middleware Assembly (PRAGMA), http://www.pragma-grid.net/
4. Distributed European Infrastructure for Supercomputing Applications (DEISA), http://www.deisa.org/
5. SETI@home, SETI@home
6. SETI@home Status: Search For Extraterrestrial Intelligence (SETI), http://www.straddle3.net/context/int/i_001218.en.html
7. Top500 List and site, http://www.top500.org
8. “The LINPACK Benchmark: Past, Present, and Future” by Jack Dongarra, Piotr Luszczek, and Antoine Petitet. Provides details of the benchmark and performance data for a number of machines. http://www.netlib.org/utk/people/JackDongarra/PAPERS/hpl.pdf
9. Gordon Bell Prize, http://www.sc2000.org/awards/
10. Supercomputing Conference Series, http://www.supercomp.org/
11. Compact Muon Solenoid (CMS) project, http://cmsinfo.cern.ch/; http://cmsinfo.cern.ch/Welcome.html/CMScollaboration/CMScollaboration.html
12. Large Hadron Collider Computing Grid (LCG), http://info.web.cern.ch/info/Press/PressReleases/Releases2003/PR13.03ELCG-1.html
13. International Virtual Observatory Alliance (IVOA), http://www.ivoa.net; National Virtual Observatory (NVO), http://www.us-vo.org/
14. Supercomputing Center of the Computer Network Information Center, Chinese Academy of Sciences, http://www.sc.ac.cn/english.htm
15. Computer Network Information Center (CNIC), Chinese Academy of Sciences (CAS), http://www.cnic.ac.cn/english/about/index.htmll
16. San Diego Supercomputer Center (SDSC), http://www.sdsc.edu/
17. National Center for Supercomputing Applications (NCSA), http://www.ncsa.uiuc.edu/
18. Pittsburgh Supercomputing Center (PSC), http://www.psc.edu/
19. Top500 list, www.top500.org, compiled by University of Tenessee, University of Manheim, NERSC/LLNL; article in People’s Daily at http://english.peopledaily.com.cn/200311/19/eng20031119_128576.shtml; Recent HPC Activities in China, Asian Technology Information Program Report ATIP03.035a 2003. http://www.atip.org/public/atip.reports.03/atip03.014.pdf
20. Pacific Rim Applications and Grid Middleware Assembly (PRAGMA), http://www.pragma-grid.net/
21. China National Grid (CNGrid), http://www.csis.hku.hk/~clwang/grid/CNGrid.html
22. China Grid Forum, http://www.chinagridforum.org/
23. GLORIAD Network, http://www.gloriad.org/
24. Asia-Pacific Advanced Network (APAN), http://www.apan.net/
25. Asia-Pacific Grid (ApGrid), http://www.apgrid.org/
26. People’s Daily, http://english.people.com.cn/200311/08/eng20031108_127892.shtml
27. The Earth Simulator Center, http://www.es.jamstec.go.jp/esc/eng/
28. Distributed European Infrastructure for Supercomputing Applications (DEISA), http://www.deisa.org/
29. Enabling Grids for E-Science in Europe (EGEE), http://egee-ei.web.cern.ch/egee-ei/
30. Japanese Grid information, http://www-1.ibm.com/grid/grid_press/pr_730.shtml
31. UK e-Science program, http://www.rcuk.ac.uk/escience/
32. “The UK e-Science Core Program and the Grid” by Tony Hey and Anne E. Trefethen, http://www.rcuk.ac.uk/escience/documents/CoreProgGrid.pdf