Lapas attēli
PDF
ePub

because they attract an interesting cross-section of users who are employing them to solve a variety of practical problems. At this meeting, even with such a small set of papers, the electric applications were obvious-data acquisition, control of power converters, VLSI logic simulation, car navigation, underwater acoustic communication, etc. There were also papers on numerical computation, including Cholesky decomposition, FFT, a two-dimensional particle-in-cell (PIC) for plasma simulation, parallel Lax-Wendroff algorithm, and a parallel implementation of 0-1 knapsack problems. For almost all of these papers, techniques already exist to solve the problems addressed or parallel algorithms are already known. The main emphasis here was to obtain an efficient implementation. For example, the FFT paper deals with an algorithm for implementing a 1-D or 2-D FFT on an eight-neighbor processor array. Such an array is obtained in software using the four communication links on each transputer. (The discrete Fourier transform is developed in powers of four, rather than two.) The linear algebra paper describes a variety of experiments on variously banded systems. There were a few papers with a computer science focus, message routing, and reduction by message passing. Two very interesting papers related to constraint satisfaction (using continuous and fuzzy variables), and multiagent planning. In these cases, the transputer is not a key ingredient, and Occam is simply used as an implementation language.

What makes transputers practical and intriguing for most of the speakers was that real parallel computing could be done with very small systems, typically only a few transputers. For example, the PIC paper deals with a problem that routinely is tasked to the largest supercomputers,

involving a large Poisson solver and many particles. Here eight transputers were employed (maximum performance possible from the hardware was about 15 MFLOPS) and a total of 1000 particles were used. The authors were disappointed with the parallelization performance because the algorithm required too much waiting time between computation. A more positive result was obtained from a parallel implementation of the modified (explicit) Lax-Wendroff method, on a twodimensional grid. The application applies to models of ionized gas in applies to models of ionized gas in the solar atmosphere. The authors have a system consisting of three boards of eight transputers each in a NEC PC plus one transputer used as a "root"; the boards are produced by Concurrent Systems. For some reason, their implementation only used eight transputers plus the root, but eight transputers plus the root, but they obtained almost linear speedup for this two-dimensional problem written in Fortran with parallel extensions. The authors conclude that while the run time for their system is about 20 times that of a Facom VP-200 supercomputer, both turnaround time and available memory are similar. Nevertheless, one feels that their primary reason for using the transputer system was that "unfortunately, we cannot access any supercomputer from our institute, so that it is difficult for us to contribute to the advance of astrophysical MHD in spite of our enthusiasm." Similarly, most of the other papers discussed systems of comparable size.

I attended this conference for two basic reasons-to get a scale of the work going on here (modest) and to hear about the activities in India.

Dr. Ashok Joshi

Center for Development of Advanced Computing (CDAC)

Poona University Campus,

Ganesh Khind

Pune 411 007 INDIA

Tel: +91 212-332461; Fax: +91 212-337551 Email:

JOSHI@PARCOM.ERNET.IN

Dr. Joshi gave an overview of CDAC. I haven't yet been there yet but am scheduled to attend a meeting in December. Thus my description is based on Joshi's remarks only.

CDAC was set up by the Indian government about four years ago to address the high performance computing requirements of the country. At that time it was recognized that there were substantial computing needs not being met. These were in standard application areas such as fluid dynamics, engineering design, computational physics and chemistry, image and signal processing, climate modeling, and biotechnology. What computational work existed was being done on very low power mainframes. A combination of political and economic factors prevents existing parallel and supercomputers from being used in India.

To address the applications needs as well as to grow research and development efforts in both hardware and software, it was decided to build and commercialize a parallel computer with peak performance around 1 GFLOP, and also to develop and support the systems and application software associated with such a computer.

After looking at various possibilities, it was decided that Inmos T800 transputers offered the easiest route for a scalable, flexible, reliable system that would be able to run a variety of applications. A system was designed by grouping the T800s in a cluster of 64 nodes. Each node consists of a T805 transputer and 4 or 16 MBYTES of memory. The cluster contains four 96 x 96 cross-point switch planes. Of the 96 links, most

are used for the compute nodes, but some are for I/O to disks, host interfaces, spares, and 16 are for interconnecting between clusters. A 256node system (four clusters) was running by mid-1991. By chaining the systems it is possible to build 1024 or larger node machines. Up to 18 hosts can be connected to a 256-node PARAM (the name is both an acronym for Parallel Machine and also Sanskrit for Supreme), which is designed to sit as a back end to a Sun or similar system. At this moment there are over 50 cluster installations within India (Joshi estimates the commercial value to be about about US$9M.)

A PARAM board containing four nodes and memory can be replaced with one that also has an i860 with 8 MB memory and a 60 MB/s data transfer mechanism between the 1860 and the transputers. The 1860s can be viewed either as vector accelerators attached to the transputers (via remote procedure calls), or the transputer's communication links can be used to develop parallel code for the 860s.

CDAC has also built a large and impressive-sounding collection of software. This includes compilers (C and Fortran,) system software, utilities, and tools. Just about everything was done ab initio; Joshi admitted that while economics was a factor, it was also done to learn the ropes and also to satisfy what he called an Indian "obsession" with doing everything themselves. In the applications area Joshi explained that there are some collaborations with Russian scientists. He commented that the well-known Russian strength in theory had overshadowed the fact that they also had real ability in computer implementation and had developed excellent user interfaces for their application software.

In addition to the basic utilities for ordinary and parallel programming, CDAC has visualization tools and a substantial collection of specific applications, including image processing, finite element analysis, digital signal processing, synthetic aperture radar analysis, computational fluid dynamics, auditory spectrograms, logic simulation, protein sequencing, and electron structure. He also emphasized that there was significant work in parallel libraries and parallel numerical algorithms, both of real interest to me. He gave only a few performance figures; one was solving a 1 K x 1 K linear system (Linpack) on a 64-node cluster using CDAC's Fortran compiler. This ran at 32 MFLOPS, about 82% of the maximum possible speedup from one node, but about half as fast as a careful Occam implementation. CDAC had also obtained about 110 MFLOPS on a 256-node machine with preconditioned conjugate gradi

ent.

One of the U.K. attendees asked if these figures weren't low compared with commercially available parallel systems. This question was well intentioned, but it seems to me its answer was essentially irrelevant. PARAM was built to eliminate the need to buy commercial systems and also to develop in-country expertise. As long as India can justify the time and labor costs of the development as opposed to the real financial cost of foreign hardware purchases, and as long as its user base is moderately satisfied with the performance obtained, this kind of activity is both sensible and prudent.

For the future CDAC is going to build a PARAM model based on the new T9000 transputer. In principle, such a system would have TFLOP performance capability. Also faster peripherals will be developed and

[blocks in formation]
[blocks in formation]

INSTITUTE FOR NEW GENERATION COMPUTER TECHNOLOGY (FIFTH GENERATION COMPUTING) 1992 (FINAL CONFERENCE)

A brief description of the Fifth Generation Computer
System, 1992 (final) conference, and an evaluation are given.

by David K. Kahaner

I have written several reports on the Fifth Generation Computer Systems (FGCS) project, and its associated research institute, Institute for New Generation Computer Technology (ICOT). See for example, [icot.kl1, 28 Feb 1992].

During the week 1 to 5 June the 1992 FGCS international conference was held, and almost 2,000 people attended. I am hoping to receive detailed reports from several attendees and will distribute these when they arrive, but in the meantime I will summarize some of the important developments very briefly.

The FGCS'92 conference was one of the best planned and executed international meetings that I have ever attended. The materials, lectures, and demonstrations have been organized and implemented with great care and preparation. The conference materials included a twovolume proceedings, complete copies of all the keynote and invited speeches, a detailed (37 page) outline of the history of the project, a complete and detailed explanation of more than twenty demonstrations, and a list and explanation of all the ICOT software. The printed material, over 2,000 pages, is in English.

It is not possible to fully summarize in this report the FGCS project. Dr. K. Fuchi, who runs the research center in Tokyo (ICOT) gave an excellent keynote speech. I was sufficiently impressed to produce a copy for inclusion with the report, below. It is both readable and interesting, and I urge you to read it.

I discussed some of my thoughts about the project with other visiting Westerners. The comments below are a quick summary of the general feelings expressed to me, as well as my own.

When FGCS was established in the early 1980s, its hopes were to use two new ideas together. First, the concept of logic programming, best known to the computing world through the languages of Prolog and Lisp, to solve very difficult computing problems of nonnumerical type. Second, parallel processing to provide the computational power to tackle the huge computational needs that logic programming would require. During the ten-year period of the project, the Japanese government spent more than 50 billion Yen (over US$300M). ICOT has had almost 200 Japanese researchers who rotate back to their home institution every

3 to 4 years. At the moment there are about 100 researchers at ICOT, almost all under 35 years old. About 75 non-Japanese researchers from 12 countries have participated in FGCS, and seven have worked at ICOT for one year or more.

At the beginning of the project, Japan's status in terms of computer science was very low, in contrast in the West, project descriptions were greeted with a combination of derision (problems are too difficult for us, no less than the Japanese, etc), and panic (if the Japanese solve these problems before we do, they will control the world's technology for decades, etc). No matter what the results, Western scientists would be able to say "we told you so". Towards the latter part of the project, the general view in the West has been that its goals have not been met. I have consistently stated in these reports, and will do so again, that my own opinion is that the impossibly high expectations claimed for FGCS were Western, and not the same as those claimed by the Japanese. The project needs to be judged against its own claims, rather than ours. Dr. Fuchi, below, makes a similar comment.

While FGCS was not another "sputnik," it had many significant accomplishments. The world has changed in ten years. The key role played by logic programming is seen by many to be reduced. Nevertheless, logic programming systems developed at ICOT are probably the best in the world. A variety of parallel computers have been built to test ideas, and some of these experimental machines are as interesting as parallel computers anywhere. The latest appears on the verge of achieving its goals of 108 logical inferences per second. Whether that speed can be attained or sustained in a real application remains to be seen. Large numbers of young Japanese have been trained in the ideas of symbolic computation, software and parallel computing, and there is probably as much or more expertise within Japanese companies on these topics as within any Western counterparts. FGCS provided both money and focus to successfully lubricate the start-up of a knowledge processing industry in Japan. The basic research from Japan in theorem proving and related areas is comparable to that in the West. A very different situation from that of ten years ago. Everyone I spoke to agreed on the major role FGCS played in the infrastructure of Japanese computer science. Also, most Western attendees were more impressed than they expected to be with the results that they heard about.

On the negative side is a lack of real applications. ICOT has developed, as the demonstrations showed, a significant number of interesting small (prototype) applications. These are running on ICOT's parallel hardware and show good speedups, which approach linear in a few cases. Perhaps ICOT should have focused on a few big applications to drive the project; these al

ways help channel and define the problems that really need to be solved rather than those that the developers are interested in solving. Perhaps ICOT hoped that industry would jump in. In fact they did not. And, given the size of the FGCS project and the number of computer makers that participated by building hardware and sending their researchers, the direct impact of FGCS on Japanese industry has been low. We repeatedly asked responsible industrial officials about the impact, and consistently got a polite "not much". It may be too early. Dr. Fuchi, below, states that we will have to wait another five years to see realistic applications. Also, ICOT developed hardware and software is unused and unusable outside of Japan, and not as much inside Japan as the ICOT people would like. A unique language and computer operating system have been developed. Presently these need to be run on special hardware, although ICOT is studying the possibility of moving the software to more standard Unix workstations.

To get more users, and also to encourage more international cooperation, MITI (Ministry of International Trade and Industry), has announced that all the ICOT developed software will be available free of charge in source form, without any restrictions as to use, modification, copying, expanding, etc. Of course no warranties are associated with such software. This amounts to over 70 large programs and includes the parallel operating software, parallel logic programming language, and all the software demonstrated at the conference.

Actually, this concept has been discussed for about a year, but MITI has now made it official. The problem of on what machines to run the programs still exists, but it is an in

telligent step by MITI nevertheless; without it the intellectual products of FGCS would likely remain unused outside of Japan. This new policy only applies to the software developed for this project. But statements by MITI about "promote the advancement of the technologies of knowledge information processing and parallel processing" suggest that it will be applied to the Real World Computing (RWC) project, MITT's new ten year project.

Concerning the demonstrations,

I should mention that these were exceptionally well done. Within ten booths, every major ICOT software product was shown, with mini-lectures, hands on, video tapes, etc., with a good combination of overview and detail. These included the following:

• Parallel inference systems, Diagnostic and control expert system based on a plant model,

• Experimental adaptive model-based diagnostic

system,

• Case-based circuit design

support system,

• Experimental parallel hierarchical recursive layout system,

• Parallel cell placement experimental system,

• High level synthesis system, Cooperative logic design expert system,

• Parallel LSI (large-scale
integration) router,

• Parallel logic simulator,
• Protein sequence analysis

program,

• Model generation theorem prover,

• Parallel database management system,

• Knowledge representation language,

« iepriekšējāTurpināt »