Outline of Achievements of RWC Project

Yoshikuni Okada Manager, Research Planning Division RWCP

I'd like to introduce the results of the recent RWC Project (Fig. 1). "Recent" means one and a half years since the previous symposium proceedings were published, so please note that I will not cover the results of the whole period of the Project.

Fig. 1

First, I will describe the objectives of RWCP researches, followed by some of the research results (Fig. 2).

Fig. 2

RWCP researches are mainly classified into 2 fields under which 46 laboratories conduct research activities (Fig. 3). The first is Real World Intelligence. This can be considered as an advanced form of what was formerly called pattern recognition, and its objective is to enhance humans' information-processing ability with the help of computers. The underlying idea is information integration technology to integrate various information technologies such as speech recognition, still image recognition, motion image recognition, and natural language processing. Real World Intelligence also involves researches on learning and self-organization technologies.

Fig. 3

Our research system consists of the Central Laboratory and Distributed Laboratory (Fig. 4).

Fig. 4

The Central Laboratory is located in the Tsukuba Mitsui Building with 5 laboratories. The Distributed Laboratory is scattered among enterprises with 41 laboratories. This number does not correspond to the number of companies involved, as some have several laboratories. 34 laboratories are in Japan and 7 are in other countries. In addition, there are many subcontractors, mainly universities, which assist with the researches at the Distributed Laboratory and Central Laboratory. This Project is closely related to the RWI Center at ETL (Electrotechnical Laboratory).

Now I'd like to proceed to the 5 domains in Real World Intelligence (Fig. 5).

Fig. 5

The first is a domain related to the so-called human interface, but is sophisticated. The second is databases - not databases as commonly understood, but databases with self-organizing capability or the ability to consolidate and handle multimodal information. The third to fifth are Robotics, Theory, and Reconfigurable Hardware.

Parallel and Distributed Technology including Seamless Parallel and Distributed Computing (Fig. 6) consists of 3 domains: Seamless Distributed Computing, Compilers, and Parallel Applications.

Fig. 6

Here I'd like to outline my views on computing trends (Fig. 7).

Fig. 7

From Minority To Majority - Only a small fraction of people used to use computers, but now the general public use them.
Networking Is Computing - Computers are not used as stand-alone devices but together with other computers, networked on a global scale.
Infoware - Evolved from hardware/software. We should take full advantage of the enormous digital archives available over networks.
Diversity - Computers no longer mean only PCs, but also cellular phones and wrist watches. The reason I mention these is that such core trendy technologies are fully considered in the RWC Project.

Fig. 8

Now I will return to the five domains. First, Multimodal Functions (Fig. 8). In short, this means multimodal human interfaces. More specifically, humans so far have used computers via keyboards for data entry. But it was considered not good enough, so the ideal human interface has been sought. For instance, computers should understand humans just as humans do, and accept any output that is generated by humans. One such system would be able to recognize not only keyboard input, but also speech, gestures, and even facial expressions (which may not be implemented soon), and respond accordingly. This is the objective of the Multimodal Human Interface.

Several laboratories are working on gestures, natural languages, 3-dimensional motions of little fingers, and sign language. Studies are also underway on answering systems for TV phones, which, in addition to simply conveying speech, also detect gestures. And image processing cameras for monitor videos, which process images as well as capture them. And facial expression recognition. Finally, as a basic technology, studies on what part of the brain perceives multimodal information are also on the menu.

Fig. 9

Let's look at the example of online gesture recognition (Fig. 9).

This system enables us to instruct computers with gestures taught in advance, each corresponding to one of the words listed to the right. Of course, the words can be changed and learned by computers.

Fig. 10

Next, the Multimodal Database Retrieval domain (Fig. 10).

A conventional relational database can handle multimedia with speech and images added, but retrieval is still based on text. The objective is to develop more advanced systems, such as Cross Mediator, tagged text, still image, motion image, and document retrieval systems.

Fig. 11

Here's an example of CrossMediator (Fig. 11).

This is a database with a single system to handle 4 media: motion images, still images, speech, and text. Multimodal retrieval can be used via video, speech, or documents.

Fig. 12

Figure 12 shows another view of CrossMediator.

Conventionally, only text was retrieved via text, but in this case motion images and other data can also be searched via text. It has a speech recognition capability through which speech pilot can be presented as well as other associated databases.

Fig. 13

Figure 13 shows a screen on which motion image is being retrieved with a motion image. Traditionally, motion images were retrieved by tagging them with keywords, but this system can retrieve motion images via motion images. Not all combinations of retrieval are supported now, but will be by the end of the Project. Its commercialization is now underway.

Fig. 14

Next, the Robotics domain (Fig. 14). This features robots actively moving around to collect and learn information, and the researches focus on analysis of positions and situations, and ambient sound recognition. Especially, ambient sound recognition, which differs from speech recognition, is unique even at the global level. A challenge here is how this capability should be applied.

Next, the Theory domain. Usually, pure theoretical research is rarely seen in any project. The RWC Project opted to focus on the theory of stochastic reasoning and other theories to support other groups in pioneering new frontier in the real world (Fig. 15).

Fig. 15

Finally, the Real World Adaptive Systems domain (Fig. 16). For real-time multimodal processing, very high real-time performance and adaptability are important. Thus, the objective is to enable systems' algorithms to be quickly executed and dynamically changed. Real World Adaptive Systems fall between control by a computer and control by specialized LSIs. Control by computer is more general but slow, while a specialized LSI offers high performance but is difficult to design and expensive without mass production. Real World Adaptive Systems are considered to complement both.

Fig. 16

The NEC Laboratory is developing many different LSI chips for reconfigurable hardware, both static and dynamic. They have two booths in at Digital Poster site (photo, P.42). Static LSI is an FPGA containing operators, with algorithmic functions mapped onto the LSI. It features well-designed operators and signal paths, as well as a C-based development environment.

Static LSI can be designed and burnt on site, ready for use. However, its functions can not be changed, while the functions of a dynamic LSI can be changed in real time, using techniques developed at the Higuchi Laboratory in the ETL RWI Center as genetic algorithms. Using registers which affect the system's behavior as genes, it reconfigures them to change functions in real time. It is used, for instance, for automatically reconfigurable neural chips, prosthetic hands developed by ETL, learning mobile robots, and high-speed compression/decompression for printing.

Matsushita Research Institute Lab. is also developing another adaptive system, which receives parallel light beams and converts them into electric signals for parallel processing by electronic circuits, whose output in turn is converted into light. By applying optical output to the next chip, they can be cascaded. This is a unique system, and the current prototype has 8x8 = 64 parallel I/O capacity, which will be enhanced to a 16x16 chip. The technology is now ready for use, requiring only specific applications.

Fig. 17

I'd now like to move on to the second field, Parallel and Distributed Computing (Fig. 17).

Seamless computing not only bridges the seams associated with traditional parallel computers, but is a totally new style of computing.

Seamless parallel distributed computers are seen as a single processor by users and networked computers can be used as a single processor. And they can be used in a heterogeneous configuration consisting of different models and different versions of computers, connected via a high-speed network. This environment is sometimes called LASN (Local Area System Network), which is a LAN environment which provides a single system.
To achieve this, a parallel operating system is required, for which the RWCP uses the SCore System, which is described later. An easy-to-use parallel compiler and parallel programming environment are also required.

Fig. 18

Figure 18 shows one image of seamless computing.

Traditional computers are accommodated in a single enclosure or multiple enclosures to perform parallel processing with short-distance interconnections proprietary to respective vendors. On the contrary, we allow scattered deployment, which can include workstations instead of desktop computers. The system to the left will shift to the scattered system to the right. High-speed interconnection over distances of 100 m is being developed.

Fig. 19

In Figure 19, the vertical axis indicates the cost, and the horizontal axis indicates the size of computer market. In general, the market for supercomputers corresponding to the top of this figure is limited, because their expensive rental cost ranges from hundreds of millions to one billion yen for just one set - four more digits to that of PC's. Therefore, users are restricted to research institutes, universities, and corporate laboratories with annual budget in such amount for computers. Since the number of institutes is limited, the market is limited. For instance, the Agency of Industrial Science and Technology has 15 scattered laboratories, so the Information Computing Center was established to install a supercomputer. Such cases lead to a limited market. In contrast to supercomputers shown at the top of this pyramid, personal computers shown at the bottom increase in number year by year. Recent statistics show that 20-30% of homes in Japan and over 50% of homes in America have a PC and the number is still increasing. To maintain this pyramid, the middle layer has to expand. The "4-digit difference" is bridged by workstation clusters shifting to PC clusters. Comparing the performance of PCs and workstations, PCs are better at whole number operations and logical operations. This is perhaps because of the tough job of LSI design - PCs can be upgraded several times a year since they sell in bulk, but this is not the case for workstations.

"Seamless" involves a shift to "heterogeneous". This does not mean simply turning PCs into PC clusters, but extending them in the horizontal direction. Thus, opposite to traditional supercomputers, PC clusters, for instance, may extend to networked server systems and multimedia-enabled ones.

As I said earlier, an operating system is required, and RWCP has introduced the SCore system software (Fig. 20).

Fig. 20

In the Language sector, Omni OpenMP and WPP were developed by the Hitachi Laboratory. They can convert sequential Fortran programs into OpenMP programs to run conventional sequential programs in parallel. The Fujitsu Laboratory is studying parallel processing with different grain sizes.

In the Network sector, the TRC Laboratory and others are developing RHiNET. The Hitachi Laboratory, along with Keio University, is developing hardware including switches as its core.

As an example of heterogeneous research, the NEC Laboratory is developing a parallel programming environment which exploits the features of both vector processors and scalar processors. Conventional numerical computation is vector-intensive, but because of great improvement recently in the speed of scalar processors, they are now studying the speed increase attained by dividing any single job into two parts - scalar processing for computing that is not suitable for vector computers, and vector processing for computing that is suitable for vector computers. The associated language and development environment are being developed, as well as tools to estimate which is faster, vector or scalar, to compute certain parts of a program.

The Fujitsu Laboratory is developing Comet, a system which seamlessly links high-speed wide area networks with LAN environments. This was used last year in Ryuichi Sakamoto's opera that brought together images and voices from Japan, Germany, and USA.

The GMD Laboratory (Germany) is developing the PROMISE system for the parallel programming environment.

Currently PC clusters run under UNIX, and the Sumitomo Metal Laboratory is attempting to make it NT-enabled.

Fig. 21

Next, I'd like to elaborate on the SCore global operating system (Fig. 21).

Compared with Beowulf, which is PC cluster software developed in the USA, SCore features faster communication. In addition, it can be applied to a wider scope of networks. For instance, it is already able to support Gigabit Ethernet, 100Base-T, and a Gigabit network called Myrinet, and support for RHiNET is planned. It also enables many heterogeneous UNIX systems to be linked.

Fig. 22

As the first step in SCore research, homogeneous clusters were developed (Fig. 22). As the second step, heterogeneous clusters are currently under development. Homogeneous clusters already have users. Heterogeneous clusters are called SCore clusters.

Fig. 23

Figure 23 shows the specifications and network configuration of SCore clusters. Three types of network can be used, and Pentium and Alpha are used as their chipsets.

As Dr. Shimada, Director of the Research Institute mentioned already, software and research results have no meaning until they are used. The SCore system is open to the public via the Internet, and is currently available only for research purposes (Fig. 24). Therefore, most of its current users are research institutes, but thanks to the Special Measures Act for Revitalizing Industry, it may be available freely soon.

Fig. 24

Its first users include RWC Laboratories such as PAPIA, PROMISE, and Toshiba Laboratory. It was then used at the University of Bonn and University of Mannheim (Germany) and LRI and Lyons University (France), as well as at the Los Alamos National Laboratory (USA). Back in Japan, Mitsubishi Electric Corporation uses SCore for its real-time simulators mainly because of its high communication speed. The Image Group of ETL also plans to use SCore.

Fig. 25

Figure 25 shows photographs taken by Dr. Ishikawa, the manager in charge of R&D of SCore, when he visited several users in Europe.

Fig. 26

University of Bonn uses a system consisting of 128 Pentium processors. Figure 26 shows photographs taken at the University of Mannheim. They use a system consisting of 8 processors.

In the Parallel Language Compiler sector, the Omni OpenMP compiler is currently being developed.

Fig. 27

OpenMP was originally developed for shared memory, simplifying parallel programming over traditional languages such as MPI. The current OpenMP is a cluster version, and will be extended to a distributed memory version for networked clusters. Currently, Omni OpenMP can be downloaded via the Internet, and it is already being used by many people. OpenMP is supported in C++ and Fortran.

Fig. 28

But ease of use must not compromise efficiency. Figure 28 compares OpenMP with multi-threaded programming. As shown, efficiency is not compromised at least for this program.

Fig. 29

Now, we will move on to optical interconnection, RHiNET-2 (Fig. 29).

Currently 1-Gigabit Ethernet is the highest, but RHiNET-2 supports 8-Gigabit optical interconnection per port. Also, an 8x8 port CMOS switch was developed. A switching board is now being developed using this switch, which supports a connection distance of 100 meters, which is sufficient for connecting between computers in offices as well as between computers and switches. It was designed for high speed with low-latency no-packet-loss communication. RHiNET-2's switch is shown in the Digital Poster site (Fig. 30 and photo).

Fig. 30

Fig. 31

These include super-high-speed optical interconnections via optical fibers, surface emitting lasers, and spatial optical information transmission. Research using optical fibers will soon be put to use. The afore-mentioned RHiNET has been developed by the Hitachi Laboratory. To carry signals fast over a short distance, fibers are bundled in parallel, and this requires skew-reduced fibers with no signal latency among fibers. The Fujikura Laboratory is developing fiber ribbons with very small skew, while the Sumiden Laboratory is studying optical wavelength multiplexing by putting light with many different wavelengths into a single fiber and increasing the density. Connecting components are being developed by Nippon Sheet Glass.

Existing light emitting lasers emit light at the edge of a chip (i.e., 1 dimension), so they are limited in the number of attached fibers. By making them 2-dimensional, several laboratories are developing surface-emitting devices using various approaches to connect many optical fibers. They may not reach the stage of practical use by the end of the Project, but they will be used for future optical interconnections. Spatial optical information transmission is being studied by Oki Electric Industry Co., Ltd. This enables us to transmit information optically chip to chip through space in substrate.

Let's move on to the Parallel Application domain (Fig. 32).

Fig. 32

One reason that applications are important is that users don't have to be aware of parallel programming. For instance, if you wished to use "Excel for supercomputers" to perform enormous calculations, you could use Excel files unmodified. In this sense, application programs for parallel computers are very important.

Fig. 33

As applications in life science, TRC is building the PAPIA system, which is a database of protein gene codes (Fig. 33).

There are 4 ways to search this database from many aspects. PAPIA runs on SCore, and the PAPIA cluster consists of Pentium-Pro chips (Fig. 34).

Fig. 34

Recently, a compact mini-cluster was built with SCore on a very small system for portability.

Fig. 35

Figure 35 shows an Internet screen displaying the PAPIA System.

Fig. 36

According to hit statistics (Fig. 36), the most accesses naturally come from Japan, followed by USA, and then by European countries including Germany, the UK, and France.

Finally, Figure 37 summarizes this lecture in one sheet. Several themes already have users, which as I noted before, is vitally important. I hope that, by the end of the Project, more themes will have users for practical use, if not commercialized. Those which have users include the SCore software system, Omni OpenMP, Cross Mediator, NEC's adaptive system, and PAPIA. Many others are still at the prototyping stage or looking for users after prototype completion, and some are expected to be perfected within the remaining 2 years. Others are basic research in nature and will be continued.

Fig. 37

Another important aspect is a mechanism for updating technologies after the Project ends. This is especially important for software, which becomes obsolete quickly as the underlying OS, applications, and processors change rapidly. We must update and improve the functions of our technologies to cope with such changes, and in the future must consider a suitable mechanism for this purpose.