rick@cs.arizona.edu (Rick Schlichting) (11/06/90)
[Dr. David Kahaner is a numerical analyst visiting Japan for two-years under the auspices of the Office of Naval Research-Far East (ONRFE). The following is the professional opinion of David Kahaner and in no way has the blessing of the US Government or any agency of it. All information is dated and of limited life time. This disclaimer should be noted on ANY attribution.] [Copies of previous reports written by Kahaner can be obtained from host cs.arizona.edu using anonymous FTP.] To: Distribution From: David Kahaner ONRFE [kahaner@xroads.cc.u-tokyo.ac.jp] H.T. Kung CMU [ht.kung@cs.cmu.edu] Re: Aspects of Parallel Computing Research in Japan--Summary. Date: 6 Nov 1990 ABSTRACT. Some aspects of parallel computing research in Japan are analyzed, based on authors' visits to a number of Japanese universities and industrial laboratories in October 1990. This portion of the report contains a summary and recommendations. INTRODUCTION. During the two week period 1-10 October 1990, the authors attended the InfoJapan '90 meeting in Tokyo, the 1990 Japan Electronics Show, also in Tokyo, and visited a number of Japanese universities and industrial laboratories, including Kyushu University, Tsukuba University, Electrotechnical Laboratory, NEC, Hitachi, Fujitsu, Sanyo, and Matsushita research Labs. Kung also had brief visits to IBM Tokyo Research Laboratory, Intel Japan and the Institute of Statistical Mathematics. The goal of this visit to Japan was to study and assess Japanese research activities in computing generally and parallel processing in particular. Kung had been to Japan several times, but not since 1986. He knew several of the researchers we visited, as some of them had previously spent time at CMU. Kahaner had been to some of the same labs and reported on these in earlier articles. Building on these contacts we have been able to construct the summary and assessment report below. Such a report can neither be complete nor static; research projects change targets, mature, and their relevance changes with respect to other work. Because of the time limitation, we were not able to visit all the projects we wanted to see. Our comments, conclusions, and recommendations are based only on those projects familiar to both of us. Others will be reported on later and may force us to alter some of the conclusions below. The report is divided into four parts. PART 1. The current part is an overall summary and assessment, with conclusions and recommendations. Some of these are based on evaluations of specific research, others on intuition and experience. A few of our comments are controversial, and sometimes the authors disagreed with each other on the conclusions, but we decided to only issue one report in order to facilitate reading. The remaining parts contain background information about the specific projects that we saw and in some cases background about the organizations that we feel help to put their research in better perspective. PART 2. Contains descriptions of supercomputing and parallel computing activities at NEC and Fujitsu. It also contains some new benchmark figures for supercomputers from these companies. A general outline of all the remaining parts is also given at the beginning of PART 2. PART 3. Deals with parallel computing at Hitachi and Matsushita, and some observations about the Japan Electronics Show 1990. PART 4. Deals with parallel computing at Kyushu and Tsukuba Universities, Electrotechnical Laboratory, Sanyo Electric, and the New Information Technology project. The following outline describes the topics that are discussed in the various parts of this report. PART 1 (this part) OUTLINE----------------------------------------------- INTRODUCTION SUMMARY RECOMMENDATIONS PART 2 OUTLINE---------------------------------------------------------- FUJITSU OVERVIEW Company profile and computer R&D activities VP2000 series supercomputer organization and performance PARALLEL PROCESSING ACTIVITIES SP (Logic Simulation Engine) AP1000 (Cellular Array Processor) RP (Routing Processor) ATM (Asynchronous Transfer Mode) Switch MISCELLANEOUS FUJITSU ACTIVITIES Neurocomputing HMET NEC SX-3 series supercomputer organization and performance Benchmark data for SX-3, VP2000, and Cray. Comments MISCELLANEOUS NEC PARALLEL PROCESSING ACTIVITIES PART 3 OUTLINE---------------------------------------------------------- HITACHI CENTRAL RESEARCH LABORATORY HDTV PARALLEL AND VECTOR PROCESSING Hyper crossbar parallel processor, H2P Parallel Inference Machine, PIM/C Josephson-Junctions Molecular Dynamics JAPAN ELECTRONICS SHOW, 1990 HDTV Flat Panel Displays MATSUSHITA ELECTRIC Company profile and computer R&D activities ADENA Parallel Processor MISCELLANEOUS ACTIVITIES HDTV Comments about Japanese industry PART 4 OUTLINE--------------------------------------------------------- KYUSHU UNIVERSITY Profile of Information Science Department Reconfigurable Parallel Processor Superscalar Processor FIFO Vector Processor Comments ELECTROTECHNICAL LABORATORY Sigma-1 Dataflow Computer and EM-4 Dataflow Comments CODA Multiprocessor NEW INFORMATION PROCESSING TECHNOLOGY Summary Comments UNIVERSITY OF TSUKUBA PAX SANYO ELECTRIC Company profile and computer R&D activities HDTV END OF OUTLINE------------------------------------------------------- SUMMARY. In Japan, parallel processing is being seriously considered by both industry and universities, but Japan is still two or three years behind the U.S. in this area. (Parallel machines here refer to distributed memory parallel machines.) The Japanese possess outstanding hardware and packaging capabilities, and definitely have advantages in hardware. They are experimenting with a substantial number of standard as well as innovative parallel architectures and interconnection networks; some have been under development and evolution for ten years or more. Most of Japanese parallel machines at present are special purpose rather than general purpose. How quickly can they catch up? The real problems in parallel processing are software, computation models and application related issues, and, we feel that the U.S. will probably keep the lead for the foreseeable future. However, using outstanding hardware strength to remove some major bottlenecks in parallel processing is a real possibility which cannot be ignored. Japanese industry's interest in parallel machines at present is probably mostly for image building (and spin-off technologies) and for internal use (to meet some special computation demands such as logic simulation of large hardware designs) rather than for possible profit from commercial sales. This is similar to their initial reasons for entering the supercomputer business several years ago. Strictly speaking, at this time there are no general purpose parallel machines commercially available from any Japanese company. Several have been "announced", such as PAX (DSV6450) from Anritsu or AP1000 from Fujitsu, but we are not aware of any deliveries. Parallel processing research at Japanese national labs and universities in general pursues niche areas related to some specific applications or computation models. Examples include Tsukuba's PAX (for quantum chromodynamics simulation), Kyoto's ADENA (for partial differential equations computation), and ETL's dataflow machines. The projects are mostly small and generally isolated from each other. Most of the projects design new hardware which is then fabricated by an industrial partner. The projects show ample determination and some of them exhibit a high-degree of creativity, but a main difficulty seems to be the meager budgets and small staff in most university departments. However, the Japanese Government seems to have recently recognized that university projects need better support and the current budget proposals attempt to address this specific issue. In the area of supercomputers, Japanese single processor supercomputers are at least as fast as those in the U.S., based mostly on their chip and packaging expertise. They follow diligently software and applications development efforts in the U.S. and do the work mostly by themselves. In the hardware area, they excel in all aspects of supercomputers ranging from 20,000 gate high-speed (70 picosecond) logic chips, to high-density multi-chip packages and to peripherals such as high transfer rate (19.6 MBytes/sec) disk arrays. There is no single area that a company will give up. Several Japanese companies have these broad base infrastructures. The Japanese industry has some special advantages in technology transfer. Most of the researchers in industrial labs we saw were young and only a small percentage have Ph.D degrees; older, more senior research staff are often moved over to development, production or management, and thus help guide corporate policy. We saw little evidence of senior researchers giving "pure" research directions, and most of the research directions appear to be from the development side with research projects jointly defined between the research and development groups. Companies seem very well informed about research, particularly in the west, and they appear to be wise about the relation of profit to research. They show real persistence in key emerging technology, but do not appear to go crazy about anything. In general we were very impressed with pervasiveness of the technology, persistence in key technology, stability of research support, awareness of research and maturity in positioning the role of research, and the success of technology transfer and product engineering laboratories. In the area of VLSI, Japanese are well known for their manufacturing capability. However, their capability in ASIC (Application Specific Integrated Circuits) is at least equally impressive. These capabilities will be of great importance in future high performance systems. On the other hand the collaboration between university researchers in Japan or between them and their industrial counterparts is not as strong as in the U.S. There are some joint research/development projects between companies and universities, but these are not common. Perhaps the large amount of research done in industry discourages inter-organization collaboration. Computer networks available to the research community are modest by U.S. standards. Very few percentages of homes have computer access and the fastest wide area networks are mostly at 48KBits/second or 64Kbits/second. This is a significant weakness in Japanese scientific communication and one that the government is trying to turn around. The persistence of research has some downside too. We have seen some anecdotal evidence of a lack of flexibility in research directions. One example is related to dataflow computers. We have also seen this phenomena in the Prolog hangup associated with the 5-th generation project and wonder if the same might not occur in fuzzy logic in due time. Western researchers seem to have less difficulty changing directions. Our problem is actually that we sometimes move too fast to a new direction without fully exploiting or understanding an older one. On the other hand we know many Japanese researchers who are very much willing to reorient their work when science, fashion, or budgets change. Research environments in Japan and U.S. are quite different. As noted above, the industrial researchers we met were all young. Kung observed that "I don't think that I know of any serious systems researchers in Japanese industrial labs who are over 35 years old." There seem to be very few career paths in Japanese industry to accommodate senior researchers. Name cards often list people as Researcher, Senior Researcher, Chief Research, Senior Chief Researcher, but most often the more senior people have substantial responsibilities in development, management, or administration. On the other hand we know many very senior staff in U.S national labs who, much like some university professors, have only their own research to be responsible for. (We wonder if this is one reason Americans are strong in research, whereas Japanese are good in making product?) Finally, both Kahaner and Kung noticed the (typically Japanese) crowded office space for researchers. A researcher will literally have no more table space left if he or she has one workstation monitor and one PC. This must have given them all the motivation to work on flat panel displays to save space. We wonder if anyone has studied the effect of community workspace versus private workspace with respect to research productivity? RECOMMENDATIONS. This is a good time for U.S. parallel machine vendors to sell products to Japan. Marketing needs to aggressively explain the advantages of using the best products as tools for further research rather than reinventing everything. Creative marketing must also recognize and adapt to the financial constraints that are imposed on Japanese university researchers. Research support for parallel computing has been strong in the U.S. The current lead that U.S. enjoys in parallel processing is a direct result of the foresight of the industry and federal funding agencies; U.S. started to support serious research efforts in this area literally many years before Japan. To keep the lead, continued support is essential. More urgently, however, is the translation of the lead into a new, sustaining capability of using parallel processing for real-world applications. Our experiences in Japan suggest a few points that we wish to emphasize. It is important to encourage research in parallel processing software and applications, and in packaging technology specifically and manufacturing in general. This represents later activities in a product development cycle, beyond the initial design of new computation models and parallel architectures. We feel that the success the Japanese have had in developing profitable advanced technology is partly due to their emphasis on the last steps in the entire process of involving a new idea or technology into a product. It is the quality of the "last 10-percent" effort that decides who ultimately benefits from the technology by being able to offer a competitive product. However, this "last 10-percent" effort demands paying attention to incremental improvements, quite different from the traditional academic research goal which encourages conceptual, paradigm shift. This calls for expanding the goal of academic research and the corresponding adjustments in research funding policies. Parallel processing becomes most significant when the machine can scale up to have a large number of processors. We need to encourage research directly addressing software and hardware issues related to very large scale parallel machines. An example is research in fault-tolerance of large scale parallel machines, an area essential to the eventual, practical use of these machines. In Japan and U.S., large parallel systems for QCD simulations already use hundreds of boards; the reliability of these systems have been a major concern. We probably can handle 1,000 board systems, but to make teraflops machines we will likely need 10,000 boards based on technology available in the foreseeable future. There is no way we will be able to keep a teraflops machine running unless fault tolerance is designed in from the beginning. An urgent agenda for the U.S. industry and scientific communities is to substantially improve their awareness of Japanese research and development activities. U.S. gets a great deal of information on technical developments in Japan, but does less well in making it available to people who really need the information, whether for research or policy. In the high-performance computing area, U.S. researchers and developers still have very limited knowledge of what is happening in Japan, especially in Japanese industry, although the situation is improving. Technical reporting (such as this, we hope) is very helpful, but it must reach the people who are actually doing or directing the technical work. Scientific liaison activities such as NSF and DoD are important, and should be active and involved, sponsoring conferences, etc. We recommend that the liason activities be expanded significantly to allow routine participation from the U.S. industry and scientific communities. ---------------END OF PART 1----------------------------------------