Суперкомпьютер ETA10: История и Разработка

1980s supercomputer

The ETA10 is a vector supercomputer designed, manufactured, and marketed by ETA Systems, a spin off division of Control Data Corporation (CDC). The ETA10 was an evolution of the CDC Cyber 205, which can trace its origins back to the CDC STAR 100, one of the first vector supercomputers to be developed. It was then referred to tentatively as the Cyber 2XX, and later as the GF 10, before it was named the ETA10. Prototypes were operational in mid 1986, and the first delivery was made in December 1986. The supercomputer was formally announced in April 1987 at an event held at its first customer installation, the Florida State University, Tallahassee's Scientific Computational Research Institute. On 17 April 1989, CDC abruptly closed ETA Systems due to ongoing financial losses, and discontinued production of the ETA10. Many of its users, such as Florida State University, negotiated Cray hardware in exchange.

CDC had a strong history of creating powerful supercomputers, starting with the CDC 6600. One of the most famous computer architects to emerge from CDC was Seymour Cray. After a disagreement with CDC management regarding the development of the CDC 8600, he went on to form his own supercomputer company, Cray Research. Meanwhile, work continued at CDC in developing a high end supercomputer, the CDC STAR 100—led by another famous architect, Neil Lincoln. Cray Research's Cray 1 vector supercomputer was successful, beating CDC's STAR 100. CDC responded with derivatives of the STAR, the Cyber 203 and 205. The Cyber 205 was moderately successful against the Cray 1's successor, the Cray X MP. It became apparent to CDC's top management that it needed to decrease the development time for the next generation computer—thus a new approach was considered for the follow on to the Cyber 205. After spinning off from CDC in September 1983, ETA set a goal of producing a supercomputer with a cycle time less than 10ns. To accomplish this, several innovations were made. Among these was the use of liquid nitrogen for cooling the CMOS based CPUs. The ETA10 successfully met the company's initial goals (10 GFLOPS), with some models achieving a cycle time of about 7 ns (143 MHz) considered rapid by mid 1980s standards. They delivered seven liquid nitrogen cooled versions and 27 smaller, air cooled versions. The CMOS circuits produced only a fraction of the heat of previous ICs. The planned 1987 follow on was supposed to be designated Cyber 250 or ETA30, as in 30 GFLOPS. ETA was eventually reincorporated back into CDC, ceasing operations on April 17, 1989.

The ETA10 series could run either ETA's EOS operating system, which was widely criticized for various problems, or a port by Lachman Associates, a software personnel firm, of UNIX System V (Release 3). While EOS suffered a reputation for poor quality, ETA's UNIX was better received by customers. EOS was preceded by and was binary executable compatible with the CDC VSOS operating system for Cyber 205. Like VSOS, EOS had demand paged virtual memory (the VS part) with 2 page sizes for improved virtual memory performance with the ETA's faster hardware pipelines. Though it had roots in the interactive Livermore Time Sharing System (LTSS), VSOS was focused as a batch oriented operating system. VSOS was not run at very many institutions and its application oriented performance, while the historic focus for supercomputing, set its features behind the times because of its limited user base. To address this feature deficiency and to make the operating system more "normal to use", the VSOS characteristics were married with UNIX characteristics in a hybrid OS. The OS was intended to be effective for both batch work that drove the hardware to its maximum or for interactive use in development from a UNIX workstation. EOS was written mainly in Cybil, a Pascal like programming language created by Control Data for its later Cyber operating systems. It was a new effort, as VSOS was implemented in IMPL, a Fortran like language created for the LTSS implementation. The command line appearance of all these systems was similar to the lineage going back to UNIVAC EXEC*8. EOS was released with early hardware deliveries and had some of the typical problems for early OS releases. Some customers delayed payment for their supercomputer installations. ETA later released a port of UNIX for the ETA 10 line, which was more quickly accepted by their customer base. However, this port started as a single processor kernel which did not transparently exploit the hardware architecture with up to 8 large application CPUs for applications.

Despite eventual adoption of UNIX, poorly developed system software remained one flaw of the ETA10 line. According to one description of the system:
Without NSF funding, the von Neumann center could be doomed. "I don’t think we can function without federal support," says Cohen. Even if the center does operate at a vastly reduced level, its machines continue to be plagued by software problems. The NSF review panel found that the ETA10 suffered a software failure once every 30 hours, and that its ability to run programs on more than one of its eight processors at any one time was poor. Although its hardware is still considered state of the art, the overall package is an "extremely immature computer system," the panel concluded. The late delivery and operating problems contributed to this demise as well as management problems. It is a mistake to believe that ETA's demise was based solely on operating system choice or existence. The Fortran compiler (ftn200) had not changed from the CDC205. This compiler retained vendor specific programming performance features (known as the Q8* subroutine calls) in an era when supercomputer users were realizing the necessity of source code portability between architectures. Additionally, the compiler optimizations were not keeping up with existing technology as shown by the Japanese supercomputer vendors as well as the newer minisupercomputer makers and competition at Cray Research. In general, computer hardware manufacturers prior and up to that period tended to be weak on software. Libraries and available commercial and non commercial applications help build an installed user base. CDC was relatively weak in this area, and some of the best operating systems that CDC provided to customers were productized versions of an OS written by Lawrence Livermore Laboratories. According to NASA, the hardware was very poorly designed, and failed to complete any acceptance tests at the Ames Research Center. This one event is considered among CDC insiders to be the downfall of ETA, which folded as a result of NASA saying no (and in a domino effect DOD, etc. ).

There were four models of the ETA10. Any of these models could be built in either single or multi processor configurations. The Models E and G were the highest performing members of the ETA10 line, and used liquid nitrogen cooling to achieve rapid cycle times. The first to be announced was the Model E, whose processors had a 10.5 ns clock cycle (approximately 95 MHz). The Model E could support up to eight processors, for a peak performance of 8.32 GFLOPS. The later Model G, whose processors had a 7 ns clock cycle (approximately 142 MHz), had a peak performance of 10.3 GFLOPS in its maximal eight processor configuration. The Models P and Q were slower, less costly, air cooled versions. The Model P, codenamed "Piper", had processors with a 24 ns clock cycle. The faster, two processor Model Q had a 19 ns clock cycle.

Between the highest performing, liquid nitrogen cooled models (ETA10 E, G, etc.) and the cheaper, air cooled models (ETA10 P, Q, etc. ), the ETA10 line spanned a 27:1 performance range. Peak performance on the top of the line models reached 10 GFLOPS. A single processor ETA10 achieved 52 MFLOPS on the LINPACK benchmark for a matrix with a size of 100 × 100.

The ETA10 was a multiprocessor system that supported up to eight CPUs. Each CPU was similar to that of a two lane Cyber 205. One of the main innovations of the ETA10 was how the CPU was implemented: the CPU was made of 250 CMOS gate array integrated circuits mounted on a 44 layer printed circuit board (PCB). Each gate array contained 20,000 gates and was fabricated using 1.25 micrometre (μm) technology that was accessible from the VHSIC program at Honeywell. In contrast, mainstream commercial technology at the time was in the 3 to 5 μm range.