\documentstyle[twoside,epsf]{article}

\input prepictex
\input pictex
\input postpictex

\pagestyle{headings}

\newfont{\fit}{cmti10}  % Text wg. baselineskip machen wir das so !
% ---------  Definitionen und Abkuerzungen  ----------------
\def\topfive{{\sc Top500}} % Small Caps !!
\def\topone{{\sc Top100}} % Small Caps !!
\def\toptwf{{\sc Top25}} % Small Caps !!
\def\linpack{{\sc Linpack}} % Small Caps !!
\def\rmax{$\fit R_{max}$}
\def\rpeak{$\fit R_{peak}$}
\def\nhalf{$\fit N_{1/2}$}
\def\nmax{$\fit N_ {max}$}
\def\flops{flop/s}

\begin{document}
\title{Developments in Japan}
\author{GEERD-R. HOFFMANN\\
European Centre for Medium-Range Weather Forecasts\\
Shinfield Park\\
Reading, Berks. RG2 9AX\\
England\\
EMAIL: Geerd.Hoffmann@ecmwf.int\\
\\
ERIC SCHNEPF\\
Siemens Nixdorf, Scientific Computing\\
Otto-Hahn-Ring 6\\
D-81739 Munich\\
Germany\\
EMAIL: Eric.Schnepf@mch.sni.de}
\label{Ch:japan}
\date{}
\maketitle

\begin{abstract}
   The supercomputer situation in Japan has changed very little since the last
   TOP 500 report (see [1]). There has been steady progress with the
   development of the series of new machines which had been announced in late
   1994 and early 1995, but they have not yet been delivered to customers.
   However, a number of important orders outside of Japan could be gained by
   the Japanese supercomputer companies against fierce competition from the US
   manufacturers. Thus the apparent isolation of the three major Japanese
   computer manufacturers to Japan seems to come to an end. In late 1995/ early
   1996, the first of the new CMOS based supercomputers will be available to
   customers both in and outside of Japan. It remains to be seen whether all
   expectations will be met.

   With regard to the market situation of supercomputers in Japan, a
   consolidation of the market shares can be seen. The two dominant
   manufacturers for Japan represented in the TOP500 list, Cray Research, Inc.
   (CRI) and Fujitsu, kept resp. increased their number of sites while Hitachi,
   NEC and Thinking Machines Corp. (TMC) - the latter not surprisingly - lost
   out. CRI will even install the first of its fully configured T9 systems in
   Japan. IBM made strong inroads with 7 new sites listed. However, SGI did not
   succeed in copying its success elsewhere in Japan: it only managed to gain
   one entry.

   With regard to the situation world-wide, Japan has again reduced its share
   of the TOP500 sites, it can only field 73 entries after 82 last year.
   However, it added 8 to the first 50 and 11 to the first 100 entries, i.e. it
   replaced a number of smaller, outdated previous supercomputers by more
   powerful, modern ones, in particular the Fujitsu VPP500. When looking at
   Rmax Gflops/s capacity installed, Japan's share slightly decreased from 27\%
   to 25.8\% of world capacity, but increased in absolute terms from 709
   Gflops/s to 1,234 Gflops/s.

   In summary, Japan kept its position world-wide as the second largest user of
   supercomputer capacity with a safe margin of 8.9\% of world-wide installed
   capacity separating it from Europe, but trailing the US by nearly 30\%.
\end{abstract}


\section{Background}

   As detailed in [1], the Japanese supercomputer scene is dominated by the
   three major domestic computer manufacturers Fujitsu, Hitachi and NEC with a
   strong presence of CRI. The three Japanese vendors started in the early 80's
   to build and market supercomputers as the top end of their mainframe
   businesses. Their determination to achieve world leadership in this field
   seems unbroken, and the long term plans are slowly coming to fruition.

   One of the authors remembers a discussion he had in 1981 with a developer of
   Fujitsu's VP-200 systems in which he was told that the Cray machines would
   be superior in the world markets for the next two generations of machines,
   but would be equal or inferior from then onwards. Considering the relevant
   machine ranges, first generation: VP-200 - Cray X-MP, second generation:
   VP-2000 - Cray Y-MP, third generation: VPP500 - C90, and fourth generation:
   VPP300 - Cray T9/ Cray T3E, the long-term prediction of the Japanese
   colleague covering nearly 15 years looks pretty close to coming true. It
   shows the long term view taken by the Japanese industry and the strategy
   employed. Initially, superior products are imported and compared with the
   local offerings, then the local market is conquered, before finally the
   external markets are won with superior products.

   At present, we are witnessing the attempted implementation of the third
   phase of the strategy. The local Japanese market is dominated by the three
   domestic supercomputer vendors with an unwavering domestic following for
   each. All three have started to penetrate the markets outside of Japan, with
   Hitachi being the last to start marketing its MPP RISC SR2001 system range
   in 1994. Outside of Japan so far in 1995, Fujitsu has won 5 orders for its
   VPP300 range, while NEC has managed to place 4 of its SX-4 systems.

2. Current commercial offerings

   The three Japanese supercomputer vendors have decided to offer different
   architectures to their customers: Fujitsu follows along the path of
   distributed memory, proprietary vector processor systems started with their
   VPP500 range; Hitachi offers MPP systems based on commodity RISC processors
   (HP variety) besides their older shared memory, proprietary vector processor
   system S3800, while NEC markets a cluster of shared memory, proprietary
   vector processor systems with up to 32 processors per cluster node.

   Fujitsu and NEC continue to build fast vector machines employing tested
   designs. Both have succeeded in building their machines using the latest .35
   micron CMOS technology, but differ with regard to the memory architecture.
   Fujitsu believes in large, distributed, relatively slow memory (SDRAM),
   while NEC uses smaller, very fast, shared memory (SSRAM) for its cluster
   nodes. Otherwise, the processors of the two competitors are remarkably
   similar with a performance of 2.2 resp. 2 GFlops/s achieved with 8 vector
   pipes.

   Hitachi follows the route developed by the MPP RISC protagonists, such as
   IBM, Intel, Meiko, TMC and more recently CRI. They use commodity RISC chips
   with their own connecting network. Since Hitachi only came to market in 1994
   with its first generation of machines, it is at a disadvantage compared to
   its competitors who by now deliver at least their second generation of
   systems. Consequently, the market penetration of Hitachi is so far
   restricted to a few development sites. However, considering the size of
   Hitachi and its probable long-term plans, the future Hitachi MPP systems may
   become formidable competitors world-wide.

3. Current market situation

   3.1. Government market

      At first, we look at the government market in Japan. The most significant
      boost for  supercomputing happened in the Japanese fiscal year '93
      (ending March '94). In that year the  government used a supplementary
      budget to stimulate the economy which resulted in 11  additional
      supercomputer sites beside the traditional university and research lab
      centres. In the fiscal year '94 there was no such special budget.

      Nevertheless, the regular budgets  allowed for some significant
      supercomputer installations. The biggest system that was installed in
      that time frame is a VPP500/80 at the Research Lab for High Energy
      Physics. Further  VPP500 systems have been installed at the Solid State
      Physics Institute of Tokyo  University, at the Computing Centre of Kyoto
      University and at the Japan Atomic Energy  Research Lab. Great attention
      was given to the success of CRI at the Fluid Dynamics Research Institute
      of Tohoku University and the Tokyo Engineering University. Two C916
      systems plus a T3D/128 MPP-system have been installed in these big,
      traditional universities.  The Hokkaido University - a loyal Hitachi
      customer - installed a Hitachi S-3800/380 system in their large scale
      computing centre. MPP systems got only a small share of the budget.
      Beside  the already mentioned T3D system, Intel succeeded to sell a
      Paragon system to the Japan  Atomic Energy Research Lab, and the
      Electrical Power Development Corporation installed a  nCube2S system.

      The fiscal year '95 is still on-going. Few decisions have already been
      made. The computing  centre of Nagoya University continues with Fujitsu
      systems and has installed a VPP500/42.  The computing centre of Tokyo
      University ordered the MPP system from Hitachi. A SR2201 system with 1024
      processors will be installed early next year. It will be interesting to
      watch whether the computer centre of Tokyo University - one of the most
      important  customers of Hitachi - will move from vector computing to the
      MPP style. This could influence the future supercomputer strategy of
      Hitachi. So far, it is not clear whether Hitachi  will announce a
      successor vector system.

      It is well known that the big Japanese universities have long lasting
      relations to a vendor. The  general purpose computing centres tend to
      stay loyal with their vendor. However, institutes of  these universities
      are more independent and procure their own systems. Often, they decide
      for  systems from other vendors than the one of the computing centre.

   3.2. General market

      There are 73 supercomputers in Japan listed in the TOP500. This
      represents a 14.6\% share of  the 500 entries - a decrease from 82 systems
      one year ago. The accumulated Rmax  performance of these 73 systems
      reaches 1.234 TFlops/s which represents 25.6\% of the  accumulated Rmax
      performance of the TOP500. These figures show that Japanese
      supercomputer installations have in average a significantly higher
      performance than sites in  other countries. This trend continues already
      for several years. In figure xxx the distribution of the number of
      systems and the accumulated Rmax performance  are listed for different
      vendors.


              Vendor       Sites       Rmax (GFlops/s)

              Convex:       1            3.306

              Cray:        10           85.926

              Fujitsu:     25          707.852

              Hitachi:      9          157.700

              IBM:          8           61.200

              Intel:        3           25.200

              NEC:         11          164.320

              Parsytec:     1            5.246

              SGI:          1            3.700

              TMC:          4           19.100

              Total:       73         1233.550

             Figure xxx: Distribution of systems to different vendors

        The market leader in Japan is still Fujitsu with 34.25\% of the number
        of systems and even 57.38\% of the accumulated Rmax performance of the
        TOP500 sites in Japan. This outstanding relation is caused by the
        success of Fujitsu's VPP500 series. The 15 VPP500 systems in Japan
        (including the Numerical Wind Tunnel system) account already for
        667.762 GFlops/s Rmax  performance.

        When we now look at the distribution into MPP, PVP and SMP systems, we
        face the difficulty that to count the VPP500 as an MPP system may lead
        to a misinterpretation of the Japanese market. The SGI system
        (installed at SGI Japan) is the only  SMP system within the Japanese
        TOP500 list. But there are many SMP systems installed in  Japan which
        just have too few processors to make it into the TOP500 list. The
        traditional  PVP systems from CRI, Fujitsu, Hitachi and NEC account for
        37 systems and 416.036  GFlops/s. The MPP systems including the VPP500
        account for 35 systems and 813.814  GFlops/s. Does that mean that MPP
        systems have achieved a breakthrough in Japan?  If we only look at the
        MPP systems which are based on standard RISC processors we count  only
        20 systems and 146.025 GFlops/s, nevertheless this is a significant
        progress compared with 1994.

        The usage of the VPP500 systems is still very similar to the usage of
        the PVP systems. The  computing centres are using the processors partly
        in throughput mode for serial vector jobs. This enables them to use
        these resources immediately very efficiently. At the same time, users
        are incrementally parallelizing their applications. Some users do it in
        the message-passing style  with PARMACS or PVM. The majority of
        parallelization work, however, is done in Fujitsu's  VPP-FORTRAN style.
        This compiler directive based method is similar to the HPF style and
        allows users to stay very close to their original vector codes.
        Consequently, the usage of the  VPP500 systems can be claimed to be
        closer to PVP than to MPP although the VPP500 is clearly a  scalable
        system with distributed memory - a main characteristic of MPP. It will
        be interesting  to watch this development in 1996 when more scalable
        vector systems, also from NEC, will  enter the list.

        3.3. Market distribution

        Another interesting aspect is the distribution of the Japanese TOP500
        systems into the  application areas. 34 systems are installed at
        research laboratories and account for 718.023  GFlops/s. 20 systems are
        installed in the academic sector and account for 309.838 GFlops/s.  The
        classified sector plays traditionally no major role in Japan with just
        1 system with 3.800  GFlops/s. The vendors have pushed 7 systems into
        this list with impressive 140.220 GFlops/s  mainly for benchmarking
        purposes.

        Let us look in more detail at the 11 TOP500 systems employed by the
        Japanese industry. A large proportion of these systems is supplied by
        CRI probably because of the wider software range available. The
        industrial systems  are smaller on average and account for 61.669
        GFlops/s. Traditionally, the automotive industry  takes the biggest
        share of supercomputers in the Japanese industry. Suzuki Motor uses a
        Hitachi S-3800/260 system, Toyota enjoys the variety of a NEC SX-3/14R,
        a Cray T94/4  and a Fujitsu VPP500/4, Honda uses a Cray T94/3, and
        finally Nippon Denso - also a  company in the automotive sector - uses
        a Cray T94/2. It is interesting to note that the first 3  Cray T90
        systems in Japan were installed in the automotive sector. Fuji Heavy
        Industry still uses the VP2600 as well as Taisei Construction. Obayashi
        Corp. continues with its NEC SX- 3/21R system. The Japanese industry
        started also to use MPP systems. An IBM SP2/72 system  is installed in
        the nuclear power industry. Mitsubishi Electric Corp. is using a Cray
        T3D system with 64 processors.

        From a European perspective it has to be recognized that only one
        European system is listed in  the TOP500 list for Japan: a Parsytec GC
        PowerPlus system with 128 processors installed at  the Japan Institute
        of Advanced Technology. In Japan the trend to buy IBM SP2 or SGI
        PowerChallenge systems in institutes and  departments is also clearly
        visible. Many systems of these types have been installed and take  over
        workload that originally was performed on supercomputers. Many of these
        systems are  not listed in the TOP500 because they have only few
        processors that result in a Rmax  performance less than the entry level
        for the TOP500.

        When one compares the situation of today with the market several years
        ago, one recognizes  that systems like the entry-level models of the
        Fujitsu VP2000, NEC SX-3 or Hitachi S-3000  series were not only used
        for typical vector applications but also for general purpose workload.
        At that time the scalar performance of these systems was also
        attractive for general purpose  applications. Today, mainly highly
        vectorized applications represent the workload on the big  vector
        computer sites. The general-purpose workload has moved to cheaper
        systems based on  standard RISC processors. This effect is one reason
        for the decrease of the number of vector  sites in the TOP500 list. On
        the other hand, the performance of the 'real' vector sites grew
        significantly. The result is that fewer vector sites represent a bigger
        share of the aggregate performance in Japan.

     4. The Japanese TOP20

     This year's Japanese TOP20 is dominated by powerful VPP500 systems. The
     leader in Japan  which also leads the world-wide TOP500 is still the
     Numerical Wind Tunnel at National  Aerospace Lab. This system with 140
     processors which is very similar to the VPP500 has been  installed in
     early '93. The fact that an almost 3 year old system leads the TOP500 list
     shows how innovative this system was. At a time when other vendors of
     vector computers continued  with PVP systems, Fujitsu took the lead with
     the vector parallel architecture which is able to  scale beyond the limits
     of traditional PVP systems. The second most powerful system in Japan  is
     the VPP500/80 at the National Lab for High Energy Physics. On number 3 we
     find the first  representative of the new CMOS based vector computer of
     NEC. The SX-4 series has also the  potential scalability to reach top
     positions in the TOP500 list. So far, only the benchmarking  system of NEC
     entered the list. Customer installations are not yet known. The positions
     4 to  10 in the Japanese TOP20 are all occupied by Fujitsu VPP500 systems.
     The latest installations  have been performed at Nagoya University and the
     National Genetics Research Lab. The high- end PVP system of Hitachi can be
     found on positions 11 to 13. The most recent installation of an S-3800/480
     is the Meteorological Agency in Japan. On positions 14 and 15 we see
     maximum configurations of NEC's SX-3, followed by two 3-processor Hitachi
     systems, a  VPP500/15 and an SX-3/3 system. On position 20 we find the
     most powerful system in Japan  that was manufactured in the US - a Cray
     C916 at the RIPS research lab in Tsukuba.

5. Conclusions and outlook

   At present, the Japanese supercomputer manufacturers are on the brink of
   world-wide market penetration. Even though the US government funded market
   will probably remain closed to them for some time to come, their offerings
   are competitive outside of the US. Their early move to CMOS technology while
   still maintaining proven proprietary vector architectures gives them a cost
   advantage over the traditional ECL vector computers and a performance
   advantage, for the time being at least, over MPP RISC systems. At last, the
   US supremacy of the supercomputer market is challenged by credible
   suppliers.

6. References

   [1] Sverre Jarp, Wolfgang Bez: Supercomputing in Japan. In: Supercomputer
       60/61. Volume XI. number 2/3. June 1995. ISSN 0168-7875. pp. 31-44.

   [PB]

7. TOP20 Supercomputers - Japan

   [to be added from list]

\end{document}
