There are also agreements with international organizations, including several sponsored by the United Nations. All of these programs are still in either the development stage or providing very early results. It will take a number of years before a significant research tradition is established, and even then the major emphasis will remain on development, improvement, and standardization. ASIAN INSTITUTE OF AIT was founded in 1959 as the South East Asia Technology Organization (SEATO) Graduate School of Engineering, to provide graduate-level civil and hydraulic engineering education not only in Thailand, but in southeast Asia generally. Ten years later it was restructured into an international institution focusing on regional problems and their engineering, scientific, and management solutions. In 1989 AIT received the Ramon Magsaysay Award for International Understanding "in recognition of its shaping a new generation of engineers and managers committed to Asia, in an atmosphere of academic excellence and regional camaraderie." A list of AIT's basic programs (in some cases leading to Ph.D. degree) gives a clear sense of AIT's mission. development and management, • Structural engineering and construction, • Telecommunications, and • Water resources engineering. In addition to these programs there are a number of outreach programs, such as in-service training, short courses, summer courses, etc., and a regional documentation center. AIT was the one location in Thailand where the staff had electronic mail, although as mentioned above it is not fully satisfactory. AIT's campus is 400 acres large, about 30 miles north of Bangkok. (I was told that the Thai government has hopes of establishing an adjacent technical park.) Currently there are about 800 students (20% women) mostly from Asia plus 200 faculty and staff. Almost half the faculty are seconded from other institutions by countries that support AIT. In 1989, faculty and staff secondments totalled more than 70 person-years, broken down as follows: Australia (person-months) China PRC It is clear that a very significant fraction of the faculty are not Thai. I was surprised to learn that several of the faculty are Vietnamese, and in fact Prof. Kanchana (CS) felt that some of her best students were often from Viet Nam or India (approximately 60 students are from each of these countries). The campus is very attractively landscaped and includes enough dormitories and houses to accommodate most of the staff and students. There is a nine-hole golf course, and naturally a cricket field, among other things. Over the years AIT's 5,500 graduates have come from 35 countries, with the largest groups (1961-1990) as follows: 24 (30) 5% 61 National government agencies 6 (20) '6% 7 10 27 136 In FY 1989-90 AIT received almost 440M Baht (about US$35M). A significant fraction of this went to support new programs, especially a new Telecommunications Division (see below), which accounted for almost US$5M, including 26M Baht from the government of Finland for the construction of a telecommunications building. (The 1990 grants for continuing education were 57M Baht, about US$4.5M.) There is a very substantial mix of grantors. United States organizations were heavy donors in the past (we were shown several large buildings that were said to have been construc-ted with U.S. grants). The U.S. government still accounts for 4.5% of AIT's cash contributions, although several other U.S. institutions and universities increase this slightly, but current large donors include Japan, Canada, and Finland. For an institution serving the needs of Asia, costs are surprisingly high; student expenses including tuition, fees, and living expenses are estimated to be over $4,100 per trimester, i.e., over US$12,000 per year. Presumably, most of AITS students are supported by either their host government or other kinds of financial aid; only about one quarter of the students are self supporting. With such a diverse student mix, it is not surprising that English is the AIT language for instruction, documents, and service. Western visitors will feel very comfortable here, and in fact the entire senior administrative staff from President to Academic Secretary are Western. In support of the educational program are laboratories and research centers, including the following: • Environmental engineering laboratories, • Regional computer center, • Remote sensing laboratory, • Computer integrated manufacturing laboratory, • Telecommunications laboratory, • Regional research, development and experimental center for projects in agricultural, food engineering, and water resources. The computer center has an IBM 3083 with about 150 terminals and PCs. There are also some CAD workstations, image processing workstations for processing satellite-sens In the CS Division, specialization is in software engineering, AI, and information technology. But there is not much time for research, and the permanent staff that arrives is well trained from good universities in Japan, Germany, Australia, United States, United Kingdom and other places and is concerned about falling behind. The research program here does not look as strong as in some of the engineering divisions and is rather theoretical. For example, Prof. Kanchana specializes in discrete algorithm analysis. In 1990, 12 journal papers were published and four contracts (expert systems, Thai sentence analyser, logic programming, and inventory system). One exception is Prof. Huynh Ngoc Phien [HNP@AIT.TH] who states that his interests are in mathematical soft ware, computer modelling, simulation, statistics, and geometry; although I was not shown any of his work. I only learned about one collaborative activity between CS and faculty in an engineering program. Given AIT's heavy emphasis on engineering and its concurrent computer needs, there are tremendous opportunities for joint research. For example, many students study structural engineering that has heavy computational demands. Similarly, the Division of Water Resources Engineering has significant needs not only in modelling but in database design and utilization. At present, computational courses are offered in specialty departments, such as "Computational Techniques in Irrigation (Differential equations, linear algebra, numerical methods, application in irrigation, mathematical programming)" in the Irrigation Engineering and Management Program, "Advanced Optimization Methods" in the Division of Industrial Engineering and Management, etc. AIT's engineering computing could also benefit from more power. But Prof. Ricardo Harboe, Chairman of the Division of Water Resources Engineering pointed out to me that by using PCs and workstations, while underpowered, will enable AIT graduates to transfer their skills more rapidly when they return to their home country, which is certainly not likely to have advanced computing facilities. Good point. AIT Division of Telecommunications This is the newest organization on the AIT campus. It is headed by Prof. A.B. Sharma Project Director Asian Institute of Technology Division of Telecommunications Bangkok 10501 Thailand Fax: +66-2-524-5730 who has been seconded from the Helsinki University of Technology in Finland. The government of Finland is also responsible for giving a large grant to AIT for establishing the Division and for constructing its building. For a number of years the Finnish International Development Agency has sponsored telecommunication training (in Finland) for people involved in a variety of projects in Asia. This was expensive and did not have an Asian focus, so after some evaluation by the UN International Telecommunications Union, it was decided to establish this program to train engineers in existing and emerging technologies as well as management techniques as applied to telecommunications. Prof. Sharma commented to me that within southeast Asia there is a widespread lack of efficient telecommunications networks, and also a massive disparity between services available in rural and metropolitan areas. Also, new technologies are appearing faster than developing countries can absorb them. The Division received its first 25 students (master's candidates) one year ago, May 1991. Five specialization areas are offered to the students. • Transmission systems, • Switching technology, • Telematics, • Network planning, • Telecom management. At this point all six faculty members are European, although Prof. Sharma plans to appoint some Asians in the near future. There are several laboratories, and facilities in clude a DX200, a digital exchange donated by the Finnish telecommunications company Nokia. This is a modern system and will permit students and researchers to study operation, maintenance, and performance evaluation of a real network, and also permits experiments with new services such as ISDN and videotelephony. So far, there is no research to speak of. I commented to Prof. Sharma that I would like to send him some of my reports, except that he didn't have an electronic mail address on his business card. He responded that he felt strongly that electronic communications such as voice, fax, and email were merging, that faxes were more common than electronic mail boxes, and that vast numbers of people already had an electronic address in the form of a telephone number. All true, but at the moment he can't receive anything from me. Another new laboratory in the Division of Industrial Engineering and Management is the CIM (Computer Integrated Manufacturing) Lab, which opened in September 1992. The first batch of ten students (from the Philippines, Bangladesh, Pakistan, and Thailand) arrived last summer. The lab has a collection of CAD/CAM software running on the IBM 3083 and IBM graphics workstations as well as microcomputer based software (MicroCADAM, AutoCAD, CADMAN B). The manufacturing section includes some production CNC machines and various training systems and computers. We did not have the opportunity to view this facility. The history and needs of Asian countries has influenced the programs at AIT. Institute documentation as well as personal conversations with faculty were replete with pointers to work associated with resourceconservation, industrial waste, soil, construction technology (foundations, concrete, etc), urban planning, environment, and land use. AIT differs from other academically oriented institutions in Asia that either use English extensively or rely on Western staff (such as Hong Kong University of Science and Technology, Institute for System Science in Singapore, or Pohang Institute of Technology in Korea, for example) in that AIT really does not expect to do world class research. Instead AIT is focusing on solving practical problems that can transcend borders within this part of the world. Because countries in the region have so many needs, the educational tasks confronting AIT are tremendous. Applied research here can have an immediate benefit on the lives of many human beings, and this could be a strong motivation for Western scientists who would like to visit and engage in joint work. Certainly, the Computer Science Division could ben-efit from senior visitors who wanted to help in the solution of engineering problems. AIT graduates pop up in a large number of positions throughout the region's science and technology infrastructure. With 800 students, even small contributions are visible. Countries and institutions that support AIT know that Asian memories are long and that repayment will probably come at an unexpected time or place. AIT is the most international institution that I have seen here in Asia, proof that some things can be constructively shared. constructively shared. In the long run though, AIT needs to build a base to allow it to have more (and more senior) Asian faculty. Also, I wonder if AIT will become more heavily Thai oriented, especially as other countries increase their own technological capabilities and see less necessity for the expense of sending students to other countries. INTERNATIONAL SYMPOSIUM ON The 19th Annual International Symposium on Computer Architecture (ISCA) was held from May 19 through 21, 1992 at Queensland, Australia. This event is one of the highlight conferences of the year and covers various aspects of computer design, including novel memory and cache organizations, new communication mechanisms for multiprocessors, simulation studies of processor architectures, and postmortems on experimental machine projects. In attendance were most of the prominent researchers in the field from academic institutions and industry. Following the conference, a series of more focused workshops were held for smaller groups of researchers. This report describes our impressions of the conference, those topics of interest to our own research, and a summary of the workshop we attended. Stephen W. Keckler and William J. Dally INTERNATION SYMPOSIUM The primary purpose of our trip was to present our paper Processor Coupling: Integrating Compile Time and Runtime Scheduling for Parallelism. In it we describe a novel processor architecture that exploits instruction level parallelism by using wide instruction words and manages interthread parallelism by allowing multiple threads of control to be active and executing simultaneously. This work is a part of the MIT MMachine that is currently being designed. The paper was well received and the forum provided useful feedback. Concerns raised by some in the audience centered on the complexity of the design (multiple function units and multiple threads on a single chip) and on the validity of our simulation results given that current compiler technology is more advanced than the compiler we used in our experiments. This particular conference had a bumper crop of papers on multithreaded architectures, with two sessions devoted to multithreading. The paper that was most applicable to our work on Processor Coupling was An Elementary Processor Architecture with Simultaneous Instruction Issuing from Multiple Threads, by Hirata, et al., of the Media Research Laboratory of Matsushita Electric Industrial Company in Japan. This paper describes an architecture in which multiple threads are simultaneously active and share use of multiple function units. However, each thread may issue at most one operation per cycle, and conflicts between threads for execution units are resolved by dynamic arbitration. Stalled operations are stored in standby stations, while the thread may continue to issue operations to other function units. One interesting result that they found was the effect of load-store bandwidth on performance. Their reported speedup for 8-thread slots increased from 3.2 to 5.8 when they doubled the number of load-store units. Although not discussed in the paper, the talk presented some additional work on multiple instruction issue by a single thread using a dynamic superscalar approach. Their organization differs from Processor Coupling primarily in that they centralize the thread state (including register files) and parcel operations out to function units. Processor coupling distributes as much state as possible to the function units and managers register coherence in software. Also of interest to our work on the M-Machine was Active Messages: A Mechanism for Integrated Communication and Computation, by von Eicken, Culler, Goldstein, and Schauser of the University of California at Berkeley. They argue that active messages, which is a communication mechanism that integrates message information into ongoing computation without requiring buffering of the message, is useful in fine grained parallel computation. They evaluate the cost of implementing active messages on several commercial and research machines, including our MIT J-Machine, concluding that a few additional architectural features would make implementing active messages more efficient. These include user-level interrupts as well as lightweight message handling such as fast context switching or a message coprocessor to handle messages without disturbing the computation processor. Two papers on adaptive routing in multicomputer networks might be of interest in the design of the MMachine network; Planar-Adaptive Routing: Low-cost Adaptive Networks for Multiprocessors by Chien and Kim from the University of Illinois, and The Tum Model for Adaptive Routing by Glass and Ni of Michigan State University. The planar-adaptive technique routes messages in a k-ary n-cube in a series of 2-D planes, thus simplifying the routing algorithm and the switch complexity. The turnmodel argues that the deadlock in adaptive networks caused by a cycle in the resource acquisition graph can be eliminated without requiring additional virtual channels; the solution prohibits particular turns (changes in dimension). This results in a family of algorithms such as "west-first" and "north-last." Both planar-adaptive routing and the turn model provide a restricted form of adaptive routing that is faster and less expensive to implement than fully adaptive. One last note about the conference concerns the invited talk on the MIPS R4000 given by John Hennessy of Stanford University and MIPS Corporation. He spent most of the talk describing the architecture, but we were most interested in his comments on the design effort. They spent two years going from design specification to first silicon and another year to reach full production (first quarter, 1992). In the process, they had a peak of 56 hardware designers (33 sustained); not included were any software people. The entire project required an investment of approximately $40 million. Dr. Hennessy's viewpoint, then, was that it is not feasible to build interesting hardware in an academic setting because of the design effort and cost required. It is interesting to contrast this with the design of the Message Driven Processor (MDP) here at MIT, which took six designers four years to go from specification to production silicon. The MDP is the processing node for the JMachine. Multiprocessor Workshop Following the conference, we attended the workshop "Future Multicomputers: Beyond Minimalist Multiprocessors?", in Hamilton Island, Australia. Two major issues were discussed: compilation and grain size of multicomputers. Monica Lam of Stanford University first discussed the compilation problem of parallel programs for multicomputers. Her thesis was that to manage parallelism effectively, a multicomputer must reduce communication and synchronization, as well as hide latencies, data coherency, and address translation mechanisms. Dr. Lam claims that this leads direct ly to coarse grained tasks and the use of prefetching remote data instead of multithreading to hide latency. However, she has assumed that the costs of communication are in fact large and did not consider fine grained tasks with fast communication primitives. Dr. Marina Chen of Yale University then described the tension between exploiting locality and ease of programming. She reiterated the importance of inexpensive communication and synchronization in developing parallel applications. The grain size discussion divided the workshop participants into "finegrained" and "coarse-grained" camps, with the latter group being much larger. Grain size was defined as the amount of memory local to a processor in a multicomputer. William Dally of MIT argued for fine grained nodes based upon silicon area costs. Given a machine with a fixed amount of memory, Dr. Dally claimed that since processor area is dwarfed by DRAM area, balancing processor and memory costs results in more nodes, each with 5-10 Mbytes of memory. The coarse grained camp demanded proof of programmability of such a machine. John Hennessy of Stanford University argued in favor of coarse grained nodes based on integrated circuit market prices. He conceded that fine grained machines have better performance per dollar, and in some restricted cases (Vector, SIMD) are easier to compile too. However, he claimed that multicomputers should be built with off-the-shelf microprocessors, which will provide an upwardly compatible product line, better developed software systems, and a large amount of memory (50-100 Megabytes per node). Dr. Dally's arguments were consistent with his view of experimental machines, while Dr. Hennessy showed his pragmatism and ties to products of the next 2-4 years. In summary, the papers presented at the International Symposium on Computer Architecture represent state-of-the-art research in computer design. More so than in the past, research is focusing on multiprocessor and parallel computer architectures, and emphasis is being placed on multithreading and on interprocessor communication. Other researchers are trying to push uniprocessor performance by exploiting instruction level parallelism through both compiler and runtime techniques. The subsequent work |