3
Brief History of Supercomputing

This chapter touches on the role, importance, and special needs of supercomputing.1It outlines the history of supercomputing, the emergence of supercomputing as a market, the entry of the Japanese supercomputing manufacturers, and the impact of supercomputing on the broader computer market and on progress in science and engineering. It focuses on hardware platforms and only touches on other supercomputing technologies, notably algorithms and software. A more detailed discussion of current supercomputing technologies is provided in Chapter 5.

THE PREHISTORY OF U.S. SUPERCOMPUTING

The development of computer technology in the United States was inextricably linked to U.S. government funding for research on cryptanalysis, nuclear weapons, and other defense applications in its first several decades.2 Arguably, the first working, modern, electronic, digital computer was the Colossus machine, put into operation at Bletchley Park,

1  

An expanded version of much of the analysis in this chapter will be found in “An Economic History of the Supercomputer Industry,” by Kenneth Flamm, 2004.

2  

In Chapter 3, “Military Roots,” of Creating the Computer: Government, Industry, and High Technology (Brookings Institution Press, 1988), Kenneth Flamm lays out the entire panorama of government-funded projects in the late 1940s and 1950s that essentially created the early U.S. computer industry. Another good but less comprehensive source ends in the very early 1950s, when high-volume production was 20 machines: N. Metropolis, J. Howlett, and Gian-Carlo Rota, A History of Computing in the Twentieth Century (Academic Press, 1980).



The National Academies | 500 Fifth St. N.W. | Washington, D.C. 20001
Copyright © National Academy of Sciences. All rights reserved.
Terms of Use and Privacy Statement



Below are the first 10 and last 10 pages of uncorrected machine-read text (when available) of this chapter, followed by the top 30 algorithmically extracted key phrases from the chapter as a whole.
Intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text on the opening pages of each chapter. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

Do not use for reproduction, copying, pasting, or reading; exclusively for search engines.

OCR for page 28
Getting up to Speed the Future of Supercomputing 3 Brief History of Supercomputing This chapter touches on the role, importance, and special needs of supercomputing.1It outlines the history of supercomputing, the emergence of supercomputing as a market, the entry of the Japanese supercomputing manufacturers, and the impact of supercomputing on the broader computer market and on progress in science and engineering. It focuses on hardware platforms and only touches on other supercomputing technologies, notably algorithms and software. A more detailed discussion of current supercomputing technologies is provided in Chapter 5. THE PREHISTORY OF U.S. SUPERCOMPUTING The development of computer technology in the United States was inextricably linked to U.S. government funding for research on cryptanalysis, nuclear weapons, and other defense applications in its first several decades.2 Arguably, the first working, modern, electronic, digital computer was the Colossus machine, put into operation at Bletchley Park, 1   An expanded version of much of the analysis in this chapter will be found in “An Economic History of the Supercomputer Industry,” by Kenneth Flamm, 2004. 2   In Chapter 3, “Military Roots,” of Creating the Computer: Government, Industry, and High Technology (Brookings Institution Press, 1988), Kenneth Flamm lays out the entire panorama of government-funded projects in the late 1940s and 1950s that essentially created the early U.S. computer industry. Another good but less comprehensive source ends in the very early 1950s, when high-volume production was 20 machines: N. Metropolis, J. Howlett, and Gian-Carlo Rota, A History of Computing in the Twentieth Century (Academic Press, 1980).

OCR for page 28
Getting up to Speed the Future of Supercomputing in the United Kingdom, in 1943. Although it was designed and employed to break a specific German cipher system, this machine was in fact a true electronic computer and could be used, in principle, on a range of problems. The existence of this machine was classified until the 1970s. U.S. personnel working with Bletchley Park during World War II played a major role in creating the early U.S. computer industry in the decade following the war. In particular, U.S. engineers at the Naval Computing Machinery Laboratory (a National Cash Register plant in Dayton, Ohio, deputized into the war effort) were building copies or improved versions of Bletchley Park electronic cryptanalysis machines, as well as computers of their own design. American engineers involved in this effort included William Norris and Howard Engstrom—Norris later founded Engineering Research Associates (ERA), then Control Data; Engstrom was later deputy director of the National Security Agency (NSA)—and Ralph Palmer who was principal technical architect of IBM’s move into electronic computers in the 1950s. Of the 55 people in the founding technical group at ERA, where Seymour Cray had his first design job in computers, 40 came from Navy communications intelligence in Washington, 5 from the Navy lab in Dayton, and 3 from the Naval Ordnance Laboratory.3 The ENIAC, built in 1945 at the University of Pennsylvania and often credited as the first functioning electronic computer, was a larger, plug-programmable computer designed to compute artillery ballistics tables.4 Ironically, it came into existence, indirectly, as a result of the code-breaking efforts of the U.S. intelligence community. The U.S. Army’s Ballistic Research Laboratory (BRL) had originally funded a ballistics computer project at National Cash Register and had turned down a competing proposal from J. Presper Eckert and John Mauchly at the University of Pennsylvania. BRL reconsidered this decision after the National Cash Register Dayton group was drafted into producing cryptanalysis machines for the Navy and finally decided to fund the ENIAC project. 3   See Flamm, 1988, pp. 36-41, 43-45. 4   As is the case for many other technologies, there has been a heated debate about who should be credited as the inventor of the first digital computer. In addition to the Colossus and the ENIAC, the following are worth mentioning: Konrad Zuse, working in Germany, built a relay-based automatic digital computer in Germany in 1939-1941. A similar system, the Automatic Sequence Controlled Calculator (ASCC), also called the Mark I, was conceived by Howard Aiken and designed and built by IBM in 1939-1944. John Vincent Atanasoff and Clifford Berry started building an electronic digital computer at Iowa State University in 1937-1942. Although the project was not completed, Atanasoff and Berry won a patent case against Eckert and Mauchly in 1973, invalidating the patent of the latter on ENIAC as the first automatic electronic computer.

OCR for page 28
Getting up to Speed the Future of Supercomputing Princeton mathematician and War Department consultant John von Neumann heard about the existence of the ENIAC project at the BRL and involved himself in the project.5 It is reported that some of the early atomic bomb calculations (in which von Neumann was involved) made use of the ENIAC even before it was formally delivered to the Army. The link between both cryptanalytical and nuclear design applications and high-performance computing goes back to the very first computers. ENIAC’s designers, Eckert and Mauchly, built the first working stored program electronic computer in the United States in 1949 (the BINAC) and delivered it to Northrop Aircraft, a defense contractor. A number of advanced machines had been built in Britain by that time—Britain was actually leading in the construction of working electronic computers in the late 1940s. A massive U.S. government investment in computer technology in the 1950s was critical to the rapid rise of U.S. companies as the undisputed leaders in the field. The second and third computers in the United States were the SEAC (built for the National Bureau of Standards, now renamed NIST) and the ERA 1101 (built for predecessors to the National Security Agency). Both went into operation in 1950, runners-up in the United States to the Eckert-Mauchly BINAC. The first Eckert and Mauchly-designed computer targeting a commercial market, the UNIVAC, was delivered to the Census Bureau in 1951. The experimental MIT Whirlwind computer, built with Navy and later Air Force funding, also went into operation in 1951. Von Neumann, who had brought British computing theoretician Alan Turing to Princeton in the 1930s and was much influenced by this contact, began work on the conceptual design of a general-purpose scientific computer for use in calculations of military interest in 1946, but a working machine was not completed until 1951. This machine was intended to be a tool for scientists and engineers doing numerical calculations of the sort needed in nuclear weapons design. Versions of the first machine installed at the Institute of Advanced Studies in Princeton, the IAS machine, were built and installed at Los Alamos (the MANIAC I) in 1952 and Oak Ridge (the ORACLE) in 1953; these were the first computers installed at the nuclear weapons laboratories.6 The nuclear weapons labs-sponsored IAS design was highly influential. But the laboratories were so pressed for computing resources before these machines were delivered that they did 5   Nancy Stern. 1981. From ENIAC to UNIVAC: An Appraisal of the Eckert-Mauchly Computers. Digital Press. 6   The Argonne National Laboratory built AVIDAC (Argonne’s Version of the Institute’s Digital Automatic Computer), which was operational prior to IAS.

OCR for page 28
Getting up to Speed the Future of Supercomputing their calculations on the SEAC at the National Bureau of Standards and ran thermonuclear calculations on the floor of the UNIVAC factory in Philadelphia. Volume computer production did not begin until 1953. In that year, the first ERA 1103 was delivered to the cryptanalysts in the intelligence community, as was the first IBM 701 Defense Calculator. Twenty ERA 1103s and 19 IBM 701s were built; all were delivered to DoD customers. NSA was the primary sponsor of high-performance computing through most of the post-1103 1950s era. It sponsored the Philco 210 and the Philco 211 and cosponsored the IBM 7030 Stretch as part of its support for the Harvest system. DoD supported the development of the IBM 7090 for use in a ballistic missile early warning system. Energy lab-sponsored computers did not play a leading role at the frontiers of high-performance computing until the late 1950s. The Atomic Energy Commission (AEC) set up a formal computer research program in 1956 and contracted with IBM for the Stretch system and with Sperry Rand (which acquired both the Eckert-Mauchly computer group and ERA in the 1950s) for the Livermore Advanced Research Computer (LARC). The cosponsorship of the Stretch system by NSA and AEC required IBM to meet the needs of two different customers (and applications) in one system. It was said that balancing those demands was an important factor in the success of IBM’s system 360. SUPERCOMPUTERS EMERGE AS A MARKET With the emergence of specific models of computers built in commercial volumes (in that era, the double digits) in the 1950s, and the dawning realization that computers were applicable to a potentially huge range of scientific and business data processing tasks, smaller and cheaper computers began to be produced in significant numbers. In the early 1950s, machines produced in volume were typically separated by less than an order of magnitude in speed. By the late 1950s, the fastest, most expensive computers were three to four orders of magnitude more powerful than the smallest models sold in large numbers. By the early 1970s, that range had widened even further, with a spread now exceeding four orders of magnitude in performance between highest performance machines and small business or scientific computers selling in volume (see Figure 3.1). In the late 1950s, the U.S. government, motivated primarily by national security needs to support intelligence and nuclear weapons applications, institutionalized its dominant role in funding the development of cutting-edge high-performance computing technology for these two sets of military applications. Arguably, the first supercomputers explicitly intended as such, designed to push an order of magnitude beyond the fast-

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.1 Early computer performance. Included in this figure are the best-performing machines according to value of installations, number of installations, and millions of operations per second (MOPS). SOURCE: Kenneth Flamm. 1988. Creating the Computer: Government, Industry, and High Technology. Washington, D.C.: Brookings Institution Press. est available commercial machines, were the IBM 7030 Stretch and Sperry Rand UNIVAC LARC, delivered in the early 1960s.7 These two machines established a pattern often observed in subsequent decades: The government-funded supercomputers were produced in very limited numbers and delivered primarily to government users. But the technology pioneered in these systems would find its way into the industrial mainstream a generation or two later in commercial systems. For example, one typical evaluation holds that “while the IBM 7030 was not considered successful, it spawned many technologies incorporated in future machines that were highly successful. The transistor logic was the basis for the IBM 7090 line of scientific computers, then the 7040 and 1400 lines. Multiprogramming, memory protection, generalized interrupts, the 7   The term “supercomputer” seems to have come into use in the 1960s, when the IBM 7030 Stretch and Control Data 6600 were delivered.

OCR for page 28
Getting up to Speed the Future of Supercomputing 8-bit byte were all concepts later incorporated in the IBM 360 line of computers as well as almost all third-generation processors and beyond. Instruction pipelining, prefetch and decoding, and memory interleaving were used in later supercomputer designs such as the IBM 360 Models 91, 95, and 195, as well as in computers from other manufacturers. These techniques are now used in most advanced microprocessors, such as the Intel Pentium and the Motorola/IBM PowerPC.”8 Similarly, LARC technologies were used in Sperry Rand’s UNIVAC III.9 Yet another feature of the supercomputer marketplace also became established over this period: a high mortality rate for the companies involved. IBM exited the supercomputer market in the mid-1970s. Sperry Rand exited the supercomputer market a few years after many of its supercomputer designers left to found the new powerhouse that came to dominate U.S. supercomputers in the 1960s—the Control Data Corporation (CDC). CONTROL DATA AND CRAY From the mid-1960s to the late 1970s, the global U.S. supercomputer industry was dominated by two U.S. companies: CDC and its offspring, Cray Research. Both companies traced their roots back to ERA, which had been absorbed by Sperry Rand in 1952. A substantial portion of this talent pool (including Seymour Cray) left to form a new company, CDC, in 1957. CDC was to become the dominant manufacturer of supercomputers from the mid-1960s through the mid-1970s. Government users, particularly the intelligence community, funded development of CDC’s first commercial offering, the CDC 1604. In 1966 CDC shipped its first full-scale supercomputer, the CDC 6600, a huge success. In addition to offering an order of magnitude jump in absolute computational capability (see Figure 3.1), it did so very cost effectively. As suggested by Figure 3.2, computing power was delivered by the 6600 at a price comparable to or lower than that of the best cost/performance in mainstream commercial machines.10 8   Historical information on the IBM 7030 is available online from the Wikipedia at <http://en.wikipedia.org/wiki/IBM_7030>. 9   See <http://en.wikipedia.org/wiki/IBM_7030>; G. Gray, “The UNIVAC III Computer,” Unisys History Newsletter 2(1) (revised 1999), <http://www.cc.gatech.edu/gvu/people/randy.carpenter/folklore/v2n1.html>. 10   The benchmarks, the performance metrics, and the cost metrics used for that figure are considerably different from those used today, but the qualitative comparison is generally accepted.

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.2 Cost/performance over time. Based on data collected by John McCallum at <http://www.jcmit.com/cpu-performance.htm>. NEC Earth Simulator cost corrected from $350 million to $500 million. Note that “normalized” MIPS (millions of instructions per second) is constructed by combining a variety of benchmarks run on these machines over this 50-year period, using scores on multiple benchmarks run on a single machine to do the normalization. At this point, there was no such thing as a commodity processor. All computer processors were custom produced. The high computational performance of the CDC 6600 at a relatively low cost was a testament to the genius of its design team. Additionally, the software tools that were provided by CDC made it possible to efficiently deliver this performance to the end user. Although the 6600 gave CDC economic success at the time, simply delivering theoretical computational power at a substantially lower price per computation was not sufficient for CDC to dominate the market. Then, as now, the availability of applications software, the availability of specialized peripherals and storage devices tailored for specific applications, and the availability of tools to assist in programming new software were just as important to many customers. The needs of the government users were different. Because the specific applications and codes they ran for defense applications were often secret, frequently were tied to special-purpose custom hardware and peripherals built in small numbers, and changed quickly over time, the avail-

OCR for page 28
Getting up to Speed the Future of Supercomputing ability of low-cost, commercially available peripherals and software were often unimportant. The defense agencies typically invested in creating the software and computing infrastructure they needed (for example, NASTRAN11 and DYNA12). When some of that software became available to commercial customers after it had been made available to the first government customers, these supercomputers became much more attractive to them. In 1972, computer designer Seymour Cray left CDC and formed a new company, Cray Research. Although CDC continued to produce high-performance computers through the remainder of the 1970s (e.g., STAR100), Cray quickly became the dominant player in the highest performance U.S. supercomputer arena.13 The Cray-1, first shipped to Los Alamos National Laboratory in 1976, set the standard for contemporary supercomputer design. The Cray-1 supported a vector architecture in which vectors of floating-point numbers could be loaded from memory into vector registers and processed in the arithmetic unit in a pipelined manner at much higher speeds than were possible for scalar operands.14 Vector processing became the cornerstone of supercomputing. Like the CDC 6600, the Cray-1 delivered massive amounts of computing power at a price competitive with the most economical computing systems of the day. Figure 3.2 shows that the cost of sustained computing power on the Cray-1 was roughly comparable to that of the cost/performance champion of the day, the Apple II microcomputer. During this period, IBM retreated from the supercomputer market, instead focusing on its fast-growing and highly profitable commercial computer systems businesses. Apart from a number of larger companies flirting with entry into the supercomputer business by building experimental machines (but never really succeeding) and several smaller com- 11   NASTRAN (NASA Structural Analysis) was originally developed at Goddard Space Flight Center and released in 1971 (see <http://www.sti.nasa.gov/tto/spinoff2002/goddard.html>). There are now several commercial implementations. 12   DYNA3D was originally developed in the 1970s at the Lawrence Livermore National Laboratory to simulate underground nuclear tests and determine the vulnerability of underground bunkers to strikes by nuclear missiles. Its successor, LS-DYNA, which simulates vehicle crashes, is commercially available. 13   CDC ultimately exited the supercomputer business in the 1980s, first spinning off its supercomputer operations in a new subsidiary, ETA, and then shutting down ETA a few years later, in 1989. 14   Vector processing first appeared in the CDC STAR100 and the Texas Instruments ASC, both announced in 1972. Much of the vector processing technology, including vectorizing compilers, originated from the Illiac IV project, developed at Illinois.

OCR for page 28
Getting up to Speed the Future of Supercomputing panies that successfully pioneered a lower-end, cost-oriented “mini-supercomputer” market niche, U.S. producers CDC and Cray dominated the global supercomputer industry in the 1970s and much of the 1980s. Although it was not widely known or documented at the time, in addition to using the systems from CDC and Cray, the defense community built special-purpose, high-performance computers. Most of these computers were used for processing radar and acoustic signals and images. These computers were often “mil-spec’ed” (designed to function in hostile environments). In general, these systems performed arithmetic operations on 16- and 32-bit data. Fast Fourier transforms and digital filters were among the most commonly used algorithms. Many of the commercial array processor companies that emerged in the late 1970s were spin-offs of these efforts. The commercial array processors, coupled with minicomputers from Digital Equipment Corporation and Data General, were often used as supercomputers. The resulting hybrid system combined a commodity host with a custom component. Unlike most other supercomputers of the period, these systems were air-cooled. The 1970s also witnessed the shipment of the first simple, single-chip computer processor (or microprocessor) by the Intel Corporation, in November 1971. By the early 1980s, this technology had matured to the point where it was possible to build simple (albeit relatively low-performance) computers capable of “serious” computing tasks. The use of low-cost, mass-produced, high-volume commodity microprocessors was to transform all segments of the computer industry. The highest performance segment of the industry, the supercomputer, was the last to be transformed by this development. ENTER JAPAN By the mid-1980s, with assistance from a substantial government-subsidized R&D program launched in the 1970s and from a history of trade and industrial policy that effectively excluded foreign competitors from Japanese markets, Japanese semiconductor producers had pushed to the technological frontier in semiconductor manufacturing. Historically, the rationale for Japanese government support in semiconductors had been to serve as a stepping-stone for creating a globally competitive computer industry, since the semiconductor divisions of the large Japanese electronics companies had also produced computers sold in a protected Japanese market. Aided by their new capabilities in semiconductors and a successful campaign to acquire key bits of IBM’s mainframe technology, by the mid-1980s Japanese computer companies were ship-

OCR for page 28
Getting up to Speed the Future of Supercomputing ping cost-effective commercial computer systems that were competitive with, and often compatible with, IBM’s mainframes.15 Thus it was that the United States viewed with some concern Japan’s announcement of two government-funded computer R&D programs in the early 1980s explicitly intended to put Japanese computer producers at the cutting edge in computer technology. One was the Fifth Generation Computer System project, which was primarily focused on artificial intelligence and logic programming. The other was the High Speed Computing System for Scientific and Technological Uses project, also called the SuperSpeed project, which focused on supercomputing technology.16 At roughly the same time, the three large Japanese electronics companies manufacturing mainframe computers began to sell supercomputers at home and abroad. The Japanese vendors provided good vectorizing compilers with their vector supercomputers. Although the Fifth Generation project ultimately would pose little threat to U.S. computer companies, it stimulated a substantial government effort in the United States to accelerate the pace of high-performance computing innovation. In the 1980s this effort, led by DARPA, funded the large Strategic Computing Initiative (SCI), which transformed the face of the U.S. supercomputer industry. The prospect of serious competition from Japanese computer companies in mainstream markets also led to a series of trade policy responses by U.S. companies and their supporters in the U.S. government (see the discussion of trade policies in Chapter 8, Box 8.1). By the 1980s, Fujitsu, Hitachi, and NEC were all shipping highly capable supercomputers competitive with Cray’s products, dominating the Japanese market and beginning to make inroads into European and American markets. The vast majority of Japanese supercomputers were sold outside the United States. There were some minimal sales to the United States in areas such as the petroleum industry but few sales to U.S. government organizations. Significant obstacles faced the sales of U.S.-made supercomputers in Japan as well. Responding to these market limitations in the 1980s, U.S. trade negotiators signed agreements with the Japanese government designed to open up government procurement in Japan to U.S. supercomputer producers. (In Japan, as in the United States, the government dominated the market for supercomputers.) In the mid-1990s, the U.S. government also supported U.S. supercomputer makers in bringing an antidumping case 15   A good reference for the survey of supercomputer development in Japan is Y. Oyanagi, 1999, “Development of Supercomputers in Japan: Hardware and Software,” Parallel Computing 25:1545-1567. 16   D.K. Kahaner. 1992. “High Performance Computing in Japan: Supercomputing.” Asian Technology Information Program. June.

OCR for page 28
Getting up to Speed the Future of Supercomputing against Japanese supercomputer makers in the U.S. market. That case ultimately forced Japanese companies out of the U.S. market until 2003, when a suspension agreement was signed. INNOVATION IN SUPERCOMPUTING While one part of the U.S. government reacted by building walls around the U.S. market, DARPA and its Strategic Computing Initiative (SCI), in concert with other government agencies and programs, took the opposite tack, attempting to stimulate a burst of innovation that would qualitatively alter the industry.17 Computing technology was regarded as the cornerstone of qualitative superiority for U.S. weapons systems. It was argued that the United States could not regain a significant qualitative lead in computing technology merely by introducing faster or cheaper computer components, since Japanese producers had clearly achieved technological parity, if not some element of superiority, in manufacturing them. Furthermore, many technologists believed that continued advances in computer capability based on merely increasing the clock rates of traditional computer processor designs were doomed to slow down as inherent physical limits to the size of semiconductor electronic components were approached. In addition, Amdahl’s law was expected to restrict increases in performance due to an increase in the number of processors used in parallel.18 The approach to stimulating innovation was to fund an intense effort to do what had not previously been done—to create a viable new architecture for massively parallel computers, some of them built around commodity processors, and to demonstrate that important applications could benefit from massive parallelism. Even if the individual processors were less efficient in delivering usable computing power, as long as the parallel architecture was sufficiently scalable, interconnecting a sufficient number 17   Investments in high-performance computing were only one area funded by the SCI, which funded over $1 billion in R&D from 1983 to 1993. There are no available data that break out this investment by technology area. Other areas were electronic components, artificial intelligence and expert systems, and large-scale prototype development of advanced military systems intended to explore new technology concepts. The committee is not aware of any objective assessment of the success and utility of the program as a whole. An excellent history of the program may be found in Alex Roland and Phillip Shiman, 2002, Strategic Computing: DARPA and the Quest for Machine Intelligence, 1983-1993, Cambridge, Mass.: MIT Press. 18   Amdahl’s law states that if a fraction of 1/s of an execution is sequential, then parallelism can reduce execution time by at most a factor of s. Conventional wisdom in the early 1980s was that for many applications of interest Amdahl’s law will restrict gains in performance from parallelism to factors of tens or low hundreds.

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.13 Mean Rmax by system type. annual growth rates in performance; hybrid systems showed the least growth in Linpack performance. Trend lines fitted to Figure 3.13 have slopes yielding annual growth rates in Rmax of 111 percent for commodity systems, 94 percent for custom systems, and 73 percent for hybrid systems.31 This is considerably faster than annual growth rates in single-processor floating-point performance shown on other benchmarks, suggesting that increases in the number of processors and improvements in the interconnect performance yielded supercomputer performance gains significantly greater than those due to component processor improvement alone for both commodity and custom systems. Hybrid system performance improvement, on the other hand, roughly tracked single-processor performance gains. Nonetheless, the economics of using much less expensive COTS microprocessors was compelling. Hybrid supercomputer systems rapidly replaced custom systems in the early 1990s. Custom supercomputer sys- 31   A regression line of the form ln Rmax = a + b Time was fit, where Time is a variable incremented by one every half year, corresponding to a new TOP500 list. Annualized trend growth rates were calculated as exp(2b)– 1.

OCR for page 28
Getting up to Speed the Future of Supercomputing tems, increasingly, were being used only in applications where software solutions making use of massively parallel hybrid systems were unsatisfactory or unavailable, or where the need for very high performance warranted a price premium. Commodity high-performance computing systems first appeared on the TOP500 list in 1997, but it was not until 2001-2002 that they began to show up in large numbers. Since 2002, their numbers have swelled, and today commodity systems account for over 60 percent of the systems on the list (see Figure 3.14). Just as hybrid systems replaced many custom systems in the late 1990s, commodity systems today appear to be displacing hybrid systems in acquisitions. A similar picture is painted by data on Rmax, which, as noted above, is probably a better proxy for systems revenues. Figure 3.15 shows how the distribution of total TOP500 system performance between these classes of systems has changed over time. Furthermore, the growing marketplace dominance of commodity supercomputer systems is not just at the low end of the market. A similar FIGURE 3.14 Share of TOP500 by system type.

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.15 Share of TOP500 Rmax by system type. pattern has also been evident in the very highest performance systems. Figure 3.16 shows how the numbers of TOP20 systems in each of these categories has changed over time. A commodity system did not appear in the top 20 highest performing systems until mid-2001. But commodity supercomputers now account for 12 of the 20 systems with the highest Linpack scores. As was true with the entire TOP500 list, custom systems were replaced by hybrid systems in the 1990s in the top 20, and the hybrid systems in turn have been replaced by commodity systems over the last 3 years. This rapid restructuring in the type of systems sold in the marketplace has had equally dramatic effects on the companies selling supercomputers. In 1993, the global HPC marketplace (with revenues again proxied by total Rmax) was still dominated by Cray, with about a third of the market, and four other U.S. companies, with about another 40 percent of the market (three of those four companies have since exited the industry). The three Japanese vector supercomputer makers accounted for another 22 percent of TOP500 performance (see Figure 3.17).

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.16 Share of top 20 machines by system type. Of the five U.S. companies with significant market share on this chart, two (Intel and Thinking Machines, second only to Cray) were building hybrid systems and three (Cray, Hewlett-Packard, and Kendall Square Research) were selling custom systems.32 The makers of traditional custom vector supercomputers (Cray and its Japanese vector competitors) have about half of the market share shown if only vector computers are considered (compare to Figure 3.12). Clearly, the HPC marketplace was undergoing a profound transformation in the early 1990s. A decade later, after the advent of hybrid systems and then of commodity high-end systems, the players have changed completely (see Figure 3.18). A company that was not even present on the list in 1993 (IBM, marketing both hybrid and commodity systems) now accounts for over half of the market, Hewlett-Packard (mainly hybrid systems) now has 32   Although some of the Thinking Machines systems counted here were using older proprietary processors, most of the Thinking Machines supercomputers on this chart were newer CM-5 machines using commodity SPARC processors.

OCR for page 28
Getting up to Speed the Future of Supercomputing FIGURE 3.17 TOP500 market share (Rmax) by company, June 1993. FIGURE 3.18 TOP500 market share (Rmax) by company, June 2004.

OCR for page 28
Getting up to Speed the Future of Supercomputing roughly the same market share as all three Japanese producers did back in 1993, and other entirely new, pure-commodity U.S. vendors in this product space (Dell, Linux Networks) are now larger than two of the three traditional Japanese supercomputer vendors. The most successful Japanese producer, NEC, has about half of the TOP500 market share it had in 1993. Cray is a shadow of its former market presence, with only 2 percent of installed capability. Two other U.S. HPC vendors (Sun and SGI), which grew significantly with the flowering of hybrid systems in the late 1990s, have ebbed with the advent of commodity systems and now have shares of the market comparable to the pure commodity supercomputer vendors and self-made systems. Over the last 15 years, extraordinary technological ferment has continuously restructured the economics of this industry and the companies surviving within its boundaries. Any policy designed to keep needed supercomputing capabilities available to U.S. government and industrial users must recognize that the technologies and companies providing these systems are living through a period of extremely rapid technological and industrial change. IMPACTS Throughout the computer age, supercomputing has played two important roles. First, it enables new and innovative approaches to scientific and engineering research, allowing scientists to solve previously unsolvable problems or to provide superior answers. Often, supercomputers have allowed scientists, engineers, and others to acquire knowledge from simulations. Simulations can replace experiments in situations where experiments are impossible, unethical, hazardous, prohibited, or too expensive; they can support theoretical experiments with systems that cannot be created in reality, in order to test the prediction of theories; and they can enhance experiments by allowing measurements that might not be possible in a real experiment. During the last decades, simulations on high-performance computers have become essential to the design of cars and airplanes, turbines and combustion engines, silicon chips or magnetic disks; they have been extensively used in support of petroleum exploration and exploitation. Accurate weather prediction would not be possible without supercomputing. According to a report by the Lawrence Berkeley National Laboratory (LBNL) for DOE, “Simulation has gained equal footing to experiments and theory in the triad of scientific process.”33 In- 33   LBNL. 2002. DOE Greenbook—Needs and Directions in High-Performance Computing for the Office of Science. Prepared for the U.S. Department of Energy. April, p. 1.

OCR for page 28
Getting up to Speed the Future of Supercomputing deed, a significant fraction of the articles published in top scientific journals in areas such as physics, chemistry, earth sciences, astrophysics, and biology, depend for their results on supercomputer simulations. The second major effect supercomputing technology has had on computing in general takes place through a spillover effect. Today’s desktop computer has the capability of the supercomputers of a decade ago. Direct Contributions Supercomputers continue to lead to major scientific contributions. Supercomputing is also critical to our national security. Supercomputing applications are discussed in detail in Chapter 4. Here the committee highlights a few of the contributions of supercomputing over the years. The importance of supercomputing has been recognized by many reports. The 1982 Lax report concluded that large-scale computing was vital to science, engineering, and technology.34 It provided several examples. Progress in oil reservoir exploitation, quantum field theory, phase transitions in materials, and the development of turbulence were all becoming possible by combining supercomputing with renormalization group techniques (p. 5). Aerodynamic design using a supercomputer resulted in the design of an airfoil with 40 percent less drag than the design using previous experimental techniques (p. 5). Supercomputers were also critical for designing nuclear power plants (p. 6). The Lax report also praised supercomputers for helping to find new phenomena through numerical experiments, such as the discovery of nonergodic behavior in the formation of solitons and the presence of strange attractors and universal features common to a large class of nonlinear systems (p. 6). As supercomputers become more powerful, new applications emerge that leverage their increased performance. Recently, supercomputer simulations have been used to understand the evolution of galaxies, the life cycle of supernovas, and the processes that lead to the formation of planets.35 Such simulations provide invaluable insight into the processes that shaped our universe and inform us of the likelihood that life-friendly planets exist. Simulations have been used to elucidate various biological mechanisms, such 34   National Science Board. 1982. Report of the Panel on Large Scale Computing in Science and Engineering. Washington, D.C., December 26 (the Lax report). 35   “Simulation May Reveal the Detailed Mechanics of Exploding Stars,” ASC/Alliances Center for Astrophysical Thermonuclear Flashes, see <http://flash.uchicago.edu/website/home/>; “Planets May Form Faster Than Scientists Thought,” Pittsburgh Supercomputer Center, see <http://www.psc.edu/publicinfo/news/2002/planets_2002-12-11.html>; J. Dubinski, R. Humble, U.-L. Pen, C. Loken, and P. Martin, 2003, “High Performance Commodity Networking in a 512-CPU Teraflops Beowulf Cluster for Computational Astrophysics,” Paper submitted to the SC2003 conference.

OCR for page 28
Getting up to Speed the Future of Supercomputing as the selective transfer of ions or water molecules through channels in cellular membranes or the behavior of various enzymes.36 Climate simulations have led to an understanding of the long-term effects of human activity on Earth’s atmosphere and have permitted scientists to explore many what-if scenarios to guide policies on global warming. We have now a much better understanding of ocean circulation and of global weather patterns such as El Niño.37 Lattice quantum chromodynamics (QCD) computations have enhanced our basic understanding of matter by exploring the standard model of particle physics.38 Box 3.1 highlights the value of having a strong supercomputing program to solve unexpected critical national problems. Codes initially developed for supercomputers have been critical for many applications, such as petroleum exploration and exploitation (three-dimensional analysis and visualization of huge amounts of seismic data and reservoir modeling), aircraft and automobile design (computational fluid mechanics codes, combustion codes), civil engineering design (finite element codes), and finance (creation of a new market in mortgage-backed securities).39 Much of the early research on supercomputers occurred in the laboratories of DOE, NASA, and other agencies. As the need for supercomputing in support of basic science became clear, the NSF supercomputing centers were initiated in 1985, partly as a response to the Lax report. Their mission has expanded over time. The centers have provided essential supercomputing resources in support of scientific research and have driven important research in software, particularly operating systems, compilers, network control, mathematical libraries, and programming languages and environments.40 Supercomputers play a critical role for the national security community according to a report for the Secretary of Defense.41 That report iden- 36   Benoit Roux and Klaus Schulten. 2004. “Computational Studies of Membrane Channels.” Structure 12 (August): 1. 37   National Energy Research Scientific Computing Center. 2002. “NERSC Helps Climate Scientists Complete First-Ever 1,000-Year Run of Nation’s Leading Climate-Change Modeling Application.” See <http://www.lbl.gov/Science-Articles/Archive/NERSC-1000-Year-climate-model.html>. 38   D. Chen, P. Chen, N.H. Christ, G. Fleming, C. Jung, A. Kahler, S. Kasow, Y. Luo, C. Malureanu, and C.Z. Sui. 1998. “3 Lattice Quantum Chromodynamics Computations.” This paper, submitted to the SC1998 conference, won the Gordon Bell Prize in the category Price-Performance. 39   NRC. 1995. Evolving the High Performance Computing and Communications Initiative to Support the Nation’s Information Infrastructure. Washington, D.C.: National Academy Press, p. 35. 40   Ibid., p. 108. 41   Office of the Secretary of Defense. 2002. Report on High Performance Computing for the National Security Community.

OCR for page 28
Getting up to Speed the Future of Supercomputing BOX 3.1 Sandia Supercomputers Aid in Analysis of Columbia Disaster Sandia National Laboratories and Lockheed Martin offered Sandia’s technical support to NASA immediately after the February 1, 2003, breakup of the space shuttle Columbia. Sandia personnel teamed with analysts from four NASA Centers to provide timely analysis and experimental results to NASA Johnson Space Center accident investigators for the purpose of either confirming or closing out the possible accident scenarios being considered by NASA. Although Sandia’s analysis capabilities had been developed in support of DOE’s stockpile stewardship program, they contained physical models appropriate to the accident environment. These models were used where they were unique within the partnership and where Sandia’s massively parallel computers and ASC code infrastructure were needed to accommodate very large and computationally intense simulations. Sandia external aerodynamics and heat transfer calculations were made for both undamaged and damaged orbiter configurations using rarefied direct simulation Monte Carlo (DSMC) codes for configurations flying at altitudes above 270,000 ft and continuum Navier-Stokes codes for altitudes below 250,000 ft. The same computational tools were used to predict jet impingement heating and pressure loads on the internal structure, as well as the heat transfer and flow through postulated damage sites into and through the wing. Navier-Stokes and DSMC predictions of heating rates were input to Sandia’s thermal analysis codes to predict the time required for thermal demise of the internal structure and for wire bundle burn-through. Experiments were conducted to obtain quasi-static and dynamic material response data on the foam, tiles, strain isolation pad, and reinforced carbon-carbon wing leading edge. These data were then used in Sandia finite element calculations of foam impacting the thermal protection tiles and wing leading edge in support of accident scenario definition and foam impact testing at Southwest Research Institute. The supercomputers at Sandia played a key role in helping NASA determine the cause of the space shuttle Columbia disaster. Sandia researchers’ analyses and experimental studies supported the position that foam debris shed from the fuel tank and impacting the orbiter wing during launch was the most probable cause of the wing damage that led to the breakup of the Columbia. NOTE: The committee thanks Robert Thomas and the Sandia National Laboratories staff for their assistance in drafting this box.

OCR for page 28
Getting up to Speed the Future of Supercomputing tified at least 10 defense applications that rely on high-performance computing (p. 22): comprehensive aerospace vehicle design, signals intelligence, operational weather/ocean forecasting, stealthy ship design, nuclear weapons stockpile stewardship, signal and image processing, the Army’s future combat system, electromagnetic weapons, geospatial intelligence, and threat weapon systems characterization. Spillover Effects Advanced computer research programs have had major payoffs in terms of technologies that enriched the computer and communication industries. As an example, the DARPA VLSI program in the 1970s had major payoffs in developing timesharing, computer networking, workstations, computer graphics, windows and mouse user interface technology, very large integrated circuit design, reduced instruction set computers, redundant arrays of inexpensive disks, parallel computing, and digital libraries.42 Today’s personal computers, e-mail, networking, data storage all reflect these advances. Many of the benefits were unanticipated. Closer to home, one can list many technologies that were initially developed for supercomputers and that, over time, migrated to mainstream architectures. For example, vector processing and multithreading, which were initially developed for supercomputers (Illiac IV/STAR100/TI ASC and CDC 6600, respectively), are now used on PC chips. Instruction pipelining and prefetch and memory interleaving appeared in early IBM supercomputers and have become universal in today’s microprocessors. In the software area, program analysis techniques such as dependence analysis and instruction scheduling, which were initially developed for supercomputer compilers, are now used in most mainstream compilers. High-performance I/O needs on supercomputers, particularly parallel machines, were one of the motivations for Redundant Array of Inexpensive Disks (RAID)43 storage, now widely used for servers. Scientific visualization was developed in large part to help scientists interpret the results of their supercomputer calculations; today, even spreadsheets can display three-dimensional data plots. Scientific software libraries such as LAPACK that were originally designed for high-performance platforms are now widely used in commercial packages running on a large range of 42   NRC. 1995. Evolving the High Performance Computing and Communications Initiative to Support the Nation’s Information Infrastructure. Washington, D.C.: National Academy Press, pp. 17-18. 43   RAID is a disk subsystem consisting of many disks that increases performance and/or provides fault tolerance.

OCR for page 28
Getting up to Speed the Future of Supercomputing platforms. In the application areas, many application packages that are routinely used in industry (e.g., NASTRAN) were initially developed for supercomputers. These technologies were developed in a complex interaction involving researchers at universities, the national laboratories, and companies. The reasons for such a spillover effect are obvious and still valid nowadays: Supercomputers are at the cutting edge of performance. In order to push performance they need to adapt new hardware and software solutions ahead of mainstream computers. And the high performance levels of supercomputers enable new applications that can be developed on capability platforms and then used on an increasingly broader set of cheaper platforms as hardware performance continues to improve.